Long-tailed Image Classification on ImageNet-LT
70.1Accuracy (Many)GLMC
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GLMCBackbone=ResNeXt502026.06 | 70.1 | 52.4 | 30.4 | 56.3 | — | |
| GLMC + SAMNBackbone=ResNeXt502026.06 | 69.7 | 56.1 | 31.1 | 57.7 | — | |
| SSDBackbone=ResNeXt-50, Scheduler=1.5x2021.09 | 66.8 | 53.1 | 35.4 | 56 | — | |
| Cross EntropyBackbone=ResNeXt-50, Scheduler=1x2021.09 | 65.9 | 37.5 | 7.7 | 44.4 | — | |
| CEBackbone=ResNeXt502026.06 | 65.9 | 37.5 | 7.7 | 44.4 | — | |
| RBLBackbone=ResNeXt502026.06 | 64.8 | 49.6 | 34.2 | 53.3 | — | |
| SSDBackbone=ResNeXt-50, Scheduler=1x2021.09 | 64.2 | 50.8 | 34.5 | 53.8 | — | |
| cRTBackbone=ResNeXt-50, Scheduler=1.5x, reproduced=true2021.09 | 64.2 | 47.7 | 27.8 | 51.3 | — | |
| IWBBackbone=ResNeXt502026.06 | 64.2 | 52.2 | 40.2 | 55.2 | — | |
| DiVEBackbone=ResNeXt502026.06 | 64.1 | 50.4 | 31.5 | 53.1 | — | |
| LWSBackbone=ResNeXt-50, Scheduler=1.5x, reproduced=true2021.09 | 63.4 | 48.6 | 32.3 | 52.1 | — | |
| EWB-FDRBackbone=ResNeXt502026.06 | 63.4 | 50 | 35.1 | 53.2 | — | |
| De-confoundBackbone=ResNeXt-50, Scheduler=1x2021.09 | 62.7 | 48.8 | 31.6 | 51.8 | — | |
| cRTBackbone=ResNeXt-50, Scheduler=1x, reproduced=true2021.09 | 62.6 | 46.9 | 27.9 | 50.3 | — | |
| WD + MaxNormBackbone=ResNeXt502026.06 | 62.5 | 50.4 | 41.5 | 53.9 | — | |
| cRTBackbone=ResNeXt-50, Scheduler=1x2021.09 | 61.8 | 46.2 | 27.4 | 49.6 | — | |
| cRTBackbone=ResNeXt502026.06 | 61.8 | 46.2 | 27.3 | 49.6 | — | |
| LWSBackbone=ResNeXt-50, Scheduler=1x, reproduced=true2021.09 | 61.1 | 48 | 31.5 | 50.7 | — | |
| MARCBackbone=ResNeXt502026.06 | 60.4 | 50.3 | 36.6 | 52.3 | — | |
| LWSBackbone=ResNeXt-50, Scheduler=1x2021.09 | 60.2 | 47.2 | 30.3 | 49.9 | — | |
| τ-normBackbone=ResNeXt502026.06 | 59.1 | 46.9 | 30.7 | 49.4 | — | |
| NCMBackbone=ResNeXt-50, Scheduler=1x2021.09 | 56.6 | 45.3 | 28.1 | 47.3 | — | |
| FedLoGe2026.06 | 46.29 | 28.01 | 15.02 | 30.52 | — | |
| FedReLabase_method=FedLoGe2026.06 | 45.43 | 30.44 | 18.02 | 31.7 | — | |
| OLTRBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 43.2 | 35.1 | 18.5 | 35.6 | — | |
| Full DatasetSelection Ratio=100%, Training Cost=35.2h2026.05 | 43.2 | 35.1 | 18.5 | 35.6 | — | |
| PODSSelection Ratio=50%, Training Cost=16.2h2026.05 | 43 | 35.5 | 20.8 | 36.3 | — | |
| OLTRBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 41.9 | 33.9 | 17.4 | — | 47.4 | |
| FedYoYo2026.06 | 41.19 | 39.42 | 31.08 | 38.15 | — | |
| Plain ModelBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 40.9 | 10.7 | 0.4 | 20.9 | — | |
| FSLwFBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 40.9 | 22.1 | 15 | 28.4 | — | |
| FSLwFBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 40.8 | 21.7 | 14.5 | — | 34.7 | |
| FedReLabase_method=FedYoYo2026.06 | 40.73 | 40.06 | 33.71 | 38.78 | — | |
| Plain ModelBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 40.1 | 10.4 | 0.4 | — | 29.5 | |
| CReFFType=Classfier-retraining2024.09 | 37.44 | 21.87 | 10.29 | 26.31 | — | |
| CCVRType=Heterogeneity-oriented FL methods2024.09 | 36.72 | 20.24 | 9.26 | 25.49 | — | |
| Focal LossBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 36.4 | 29.9 | 16 | 30.5 | — | |
| Ratio LossType=Imbalance-oriented FL methods2024.09 | 36.33 | 18.14 | 7.41 | 24.31 | — | |
| Lifted LossBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 35.8 | 30.4 | 17.9 | 30.8 | — | |
| Range LossBackbone=ResNet-10, Evaluation Setting=Closed-set2019.04 | 35.8 | 30.3 | 17.6 | 30.7 | — | |
| Range Loss + OpenMaxBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 35.8 | 30.3 | 17.6 | — | 36.8 | |
| Focal LossBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 35.7 | 29.3 | 15.6 | — | 37.1 | |
| FedAvgType=Heterogeneity-oriented FL methods2024.09 | 34.92 | 19.18 | 7.1 | 23.85 | — | |
| Lifted LossBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 34.8 | 29.3 | 17.4 | — | 37.4 | |
| Range LossBackbone=ResNet-10, Evaluation Setting=Open-set2019.04 | 34.7 | 29.4 | 17.2 | — | 37.3 | |
| FedProxType=Heterogeneity-oriented FL methods2024.09 | 34.25 | 17.06 | 6.37 | 22.99 | — | |
| FedAvgMType=Heterogeneity-oriented FL methods2024.09 | 33.93 | 18.55 | 6.73 | 22.57 | — | |
| FedDFType=Heterogeneity-oriented FL methods2024.09 | 31.78 | 15.52 | 4.48 | 21.63 | — | |
| Fed-Focal LossType=Imbalance-oriented FL methods2024.09 | 31.74 | 15.77 | 5.52 | 21.6 | — | |
| FedAvg+ τ -normType=Imbalance-oriented FL methods2024.09 | 31.66 | 15.76 | 4.33 | 21.58 | — | |
| MLLM-LLaVA-FLType=Our framework2024.09 | 30.85 | 25.89 | 25.58 | 27.53 | — | |
| AREABackbone=ResNeXt502026.06 | — | — | — | 49.5 | — | |
| LogitAdjustBackbone=ResNeXt502026.06 | — | — | — | 51.1 | — | |
| OLTRBackbone=ResNeXt-50, Scheduler=1x2021.09 | — | — | — | 46.3 | — |