Multi-Label Classification on NUS-WIDE (test)
70.4mAPPanCAN
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PanCANBackbone Network=CvT-W24, Cells=8x102025.12 | 70.4 | 69.9 | 77.8 | |
| Q2L-CvTBackbone Network=CvT-W242025.12 | 70.1 | 67.6 | 76.3 | |
| PanCANBackbone Network=CvT-W24, Cells=4x52025.12 | 69.8 | 69.6 | 77.5 | |
| MCDKNBackbone Network=CvT-W24, Cells=8x102025.12 | 69.7 | 68.9 | 76.6 | |
| MCDKNBackbone Network=CvT-W24, Cells=4x52025.12 | 69.4 | 68.2 | 76.1 | |
| ML-VPTBackbone=DINOv2/B, Resolution=224x2242025.04 | 68.7 | 65.5 | 75.2 | |
| PanCANBackbone Network=TResNetL, Cells=8x102025.12 | 68.3 | 65.7 | 77.6 | |
| ML-VPTBackbone=ViT-B-21k, Resolution=224x2242025.04 | 68.2 | 65 | 75 | |
| VPTBackbone=DINOv2/B, Resolution=224x2242025.04 | 68.1 | 65.1 | 75 | |
| E2VPTBackbone=DINOv2/B, Resolution=224x2242025.04 | 67.9 | 65 | 75 | |
| MCDKNBackbone Network=TResNetL, Cells=8x102025.12 | 67.8 | 65.1 | 76.5 | |
| E2VPTBackbone=ViT-B-21k, Resolution=224x2242025.04 | 67.7 | 64.9 | 74.7 | |
| PanCANBackbone Network=TResNetL, Cells=4x52025.12 | 67.6 | 65.1 | 76.9 | |
| VPTBackbone=ViT-B-21k, Resolution=224x2242025.04 | 67.5 | 64.8 | 74.7 | |
| GateVPTBackbone=DINOv2/B, Resolution=224x2242025.04 | 67.2 | 64.6 | 74.5 | |
| GateVPTBackbone=ViT-B-21k, Resolution=224x2242025.04 | 66.9 | 64.4 | 74.3 | |
| ML-VPTBackbone=DINOv2/S, Resolution=224x2242025.04 | 66.9 | 64.3 | 74.8 | |
| PanCANBackbone Network=ResNet101, Cells=8x102025.12 | 66.9 | 65.2 | 75.5 | |
| MCDKNBackbone Network=TResNetL, Cells=4x52025.12 | 66.9 | 64.5 | 75.8 | |
| MITr-l2021.06 | 66.3 | 65 | 75.8 | |
| MCDKNBackbone Network=ResNet101, Cells=8x102025.12 | 66.3 | 64.6 | 74.8 | |
| Q2L-TResLBackbone Network=TResNetL2025.12 | 66.3 | 64 | 75 | |
| MlTr-lBackbone Network=MlTr-l(22k)2025.12 | 66.3 | 65 | 75.8 | |
| PanCANBackbone Network=ResNet101, Cells=4x52025.12 | 66.1 | 64.5 | 74.9 | |
| SADCLBackbone Network=ResNet1012025.12 | 65.9 | 63 | 75 | |
| ML-VPTBackbone=ViT-B, Resolution=224x2242025.04 | 65.7 | 63.3 | 74.1 | |
| E2VPTBackbone=DINOv2/S, Resolution=224x2242025.04 | 65.4 | 63.1 | 74.3 | |
| MCDKNBackbone Network=ResNet101, Cells=4x52025.12 | 65.4 | 63.9 | 74.2 | |
| VPTBackbone=DINOv2/S, Resolution=224x2242025.04 | 65.3 | 63.2 | 74.3 | |
| VPTBackbone=ViT-B, Resolution=224x2242025.04 | 65.2 | 62.8 | 73.8 | |
| Tresnet-l2021.06 | 65.2 | 63.6 | 75 | |
| ASLBackbone=TResNet-L2020.09 | 65.2 | 63.6 | 75 | |
| ASLBackbone Network=ResNet1012025.12 | 65.2 | 63.6 | 75 | |
| ASLBackbone Network=TResNetL2025.12 | 65.2 | 63.6 | 75 | |
| Q2L-R101Backbone Network=ResNet1012025.12 | 65 | 63.1 | 75 | |
| ML-SGMBackbone Network=ResNet1012025.12 | 64.6 | 62.4 | 72.5 | |
| GateVPTBackbone=DINOv2/S, Resolution=224x2242025.04 | 64.5 | 62.6 | 73.7 | |
| Focal lossBackbone Network=TResNetL2025.12 | 64 | 62.9 | 74.7 | |
| E2VPTBackbone=ViT-B, Resolution=224x2242025.04 | 63.9 | 62.2 | 73.7 | |
| ASLBackbone=ResNet1012020.09 | 63.9 | 62.7 | 74.6 | |
| SSTBackbone Network=ResNet1012025.12 | 63.5 | 59.6 | 73.2 | |
| ML-VPTBackbone=MoCo v3, Resolution=224x2242025.04 | 63 | 61.1 | 73.6 | |
| GateVPTBackbone=ViT-B, Resolution=224x2242025.04 | 62.8 | 61.5 | 72.9 | |
| E2VPTBackbone=MoCo v3, Resolution=224x2242025.04 | 62.8 | 61.2 | 73.5 | |
| ICME2021.06 | 62.8 | 60.7 | 74.1 | |
| ICME2020.09 | 62.8 | 60.7 | 74.1 | |
| ICMEBackbone Network=ResNet1012025.12 | 62.8 | 60.7 | 74.1 | |
| VPTBackbone=MoCo v3, Resolution=224x2242025.04 | 62.7 | 61.3 | 73.8 | |
| SRN2021.06 | 62 | 58.5 | 73.4 | |
| SRN2020.09 | 62 | 58.5 | 73.4 | |
| SRNBackbone Network=ResNet1012025.12 | 62 | 58.5 | 73.4 | |
| ML-VPTBackbone=MAE, Resolution=224x2242025.04 | 61.9 | 60.7 | 73.5 | |
| GateVPTBackbone=MoCo v3, Resolution=224x2242025.04 | 61.8 | 60.5 | 72.9 | |
| MS-CMAEvaluation labels=All2019.12 | 61.4 | 60.5 | 73.8 | |
| MS-CMAEvaluation labels=Top-32019.12 | 61.4 | 55.7 | 69.5 | |
| MS-CMA2021.06 | 61.4 | 60.5 | 73.8 | |
| MS-CMA2020.09 | 61.4 | 60.5 | 73.8 | |
| MS-CMABackbone Network=ResNet1012025.12 | 61.4 | 60.5 | 73.8 | |
| Full-tuningTrainable Parameters (M)=85.862022.05 | 61.26 | — | — | |
| CMAEvaluation labels=All2019.12 | 60.8 | 60.4 | 73.7 | |
| CMAEvaluation labels=Top-32019.12 | 60.8 | 55.5 | 70 | |
| E2VPTBackbone=MAE, Resolution=224x2242025.04 | 60.6 | 59.4 | 73 | |
| VPTBackbone=MAE, Resolution=224x2242025.04 | 60.3 | 59.3 | 73 | |
| S-CLs (2018)Evaluation labels=All2019.12 | 60.1 | 58.7 | 73.7 | |
| S-CLs (2018)Evaluation labels=Top-32019.12 | 60.1 | 53.8 | 71.1 | |
| S-CLs2020.09 | 60.1 | 58.7 | 73.7 | |
| AdaptFormer-64Trainable Parameters (M)=1.25, bottleneck_dimension=642022.05 | 59.07 | — | — | |
| FedMPTVenue=Ours, Heterogeneity ratio (t)=10%2026.05 | 58.36 | 51.56 | 78.98 | |
| GateVPTBackbone=MAE, Resolution=224x2242025.04 | 58.2 | 57.6 | 71.9 | |
| AdaptFormer-4Trainable Parameters (M)=0.15, bottleneck_dimension=42022.05 | 58.14 | — | — | |
| Attention transfer (2016)Evaluation labels=All2019.12 | 57.6 | 55.2 | 70.3 | |
| Attention transfer (2016)Evaluation labels=Top-32019.12 | 57.6 | 51.7 | 68.8 | |
| AdaptFormer-1Trainable Parameters (M)=0.09, bottleneck_dimension=12022.05 | 57.51 | — | — | |
| FitsNet (2014)Evaluation labels=All2019.12 | 57.4 | 54.9 | 70.4 | |
| FitsNet (2014)Evaluation labels=Top-32019.12 | 57.4 | 51.4 | 68.6 | |
| VPTTrainable Parameters (M)=0.072022.05 | 57.08 | — | — | |
| FedMPTVenue=Ours, Heterogeneity ratio (t)=Avg2026.05 | 56.69 | 47.77 | 77.33 | |
| CNN-RNN (2016)Evaluation labels=Top-32019.12 | 56.1 | 34.7 | 55.2 | |
| Full labelTraining Strategy=LinearInit.2022.06 | 54.9 | — | — | |
| Full labelTraining Strategy=End-to-end2022.06 | 54.5 | — | — | |
| Fed-RAMVenue=CVPR’25, Heterogeneity ratio (t)=Avg2026.05 | 53.33 | 44.33 | 74.09 | |
| FedAWAVenue=CVPR’25, Heterogeneity ratio (t)=Avg2026.05 | 52.55 | 44.04 | 70.23 | |
| FedMVPVenue=ICCV’25, Heterogeneity ratio (t)=Avg2026.05 | 52.3 | 44.39 | 75.42 | |
| Fed-MaPLeVenue=CVPR’23, Heterogeneity ratio (t)=Avg2026.05 | 52.27 | 42.93 | 72.1 | |
| LinearTrainable Parameters (M)=0.062022.05 | 51.19 | — | — | |
| ROLETraining Strategy=LinearInit.2022.06 | 51 | — | — | |
| FedTPGVenue=ICLR’24, Heterogeneity ratio (t)=Avg2026.05 | 50.93 | 45.95 | 69.73 | |
| FedPGPVenue=ICML’24, Heterogeneity ratio (t)=Avg2026.05 | 50.73 | 43.82 | 73.46 | |
| LSANTraining Strategy=LinearInit.2022.06 | 50.5 | — | — | |
| Fed-TCPVenue=CVPR’24, Heterogeneity ratio (t)=Avg2026.05 | 50.28 | 44.2 | 67.67 | |
| LL-CtTraining Strategy=LinearInit.2022.06 | 49.6 | — | — | |
| LL-CpTraining Strategy=LinearInit.2022.06 | 49.4 | — | — | |
| LL-RTraining Strategy=LinearInit.2022.06 | 49.1 | — | — | |
| LL-CpTraining Strategy=End-to-end2022.06 | 48.3 | — | — | |
| EPRTraining Strategy=LinearInit.2022.06 | 48.1 | — | — | |
| LL-CtTraining Strategy=End-to-end2022.06 | 48 | — | — | |
| CAProBackbone=R502023.10 | 48 | 39.3 | 45.4 | |
| Naive ANTraining Strategy=LinearInit.2022.06 | 47.6 | — | — | |
| WANTraining Strategy=LinearInit.2022.06 | 47.5 | — | — | |
| LL-RTraining Strategy=End-to-end2022.06 | 47.4 | — | — |