Image Classification on ImageNet-LT (val)
78.7Top-1 Acc (Total)Ours-CVS+TTE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ours-CVS+TTELoss-ori.=true, Training Protocol=Fine-tuning pre-trained model2023.10 | 78.7 | 81.2 | 78.1 | 74.2 | |
| LIFT+TTELoss-ori.=false, Training Protocol=Fine-tuning pre-trained model2023.10 | 78.3 | 81.3 | 77.4 | 73.4 | |
| Ours-CVSLoss-ori.=true, Training Protocol=Fine-tuning pre-trained model2023.10 | 77.4 | 80.1 | 76.9 | 72.4 | |
| LIFTLoss-ori.=false, Training Protocol=Fine-tuning pre-trained model2023.10 | 77 | 80.2 | 76.1 | 71.5 | |
| DecoderLoss-ori.=false, Training Protocol=Fine-tuning pre-trained model2023.10 | 73.2 | — | — | — | |
| RIDE teacherBackbone=ResNeXt-502021.03 | 57.5 | — | — | — | |
| RIDE-DIVEBackbone=ResNeXt-502021.03 | 57.12 | — | — | — | |
| RIDEBackbone=ResNeXt-502021.03 | 56.8 | — | — | — | |
| BCLCUDA=true2023.02 | 56.3 | 66.8 | 53.9 | 36.6 | |
| RIDE (3 experts) + CMOBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 56.2 | 66.4 | 53.9 | 35.6 | |
| BCLCUDA=false2023.02 | 55.6 | 65.3 | 53.5 | 36.3 | |
| RIDE-DIVE teacherBackbone=ResNeXt-502021.03 | 55.6 | — | — | — | |
| RIDE (4 experts)Backbone=ResNet-50, Training Epochs=90 or 1002021.12 | 55.4 | 66.2 | 52.3 | 36.5 | |
| RIDE (3 experts)Backbone=ResNet-50, Training Epochs=90 or 1002021.12 | 54.9 | 66.2 | 51.7 | 34.9 | |
| RIDEexperts=3, CUDA=true2023.02 | 54.7 | 65.9 | 51.7 | 34.9 | |
| Ours-CVSLoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 54.7 | 63 | 53.2 | 38.9 | |
| Ours-OriginLoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 54.1 | 62.9 | 52.6 | 37.1 | |
| RIDE + CMOCUDA=false2023.02 | 54 | 65.6 | 50.6 | 34.8 | |
| WBLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 53.9 | 62.5 | 50.4 | 41.5 | |
| DLLALoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 53.7 | — | — | — | |
| RIDEexperts=3, CUDA=false2023.02 | 53.6 | 64.8 | 50.8 | 34.6 | |
| DRO-LTLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 53.5 | 64 | 49.8 | 33.1 | |
| SBCLLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 53.4 | 63.8 | 51.3 | 31.2 | |
| DiVELoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 53.1 | 64.1 | 50.4 | 31.5 | |
| DisAlignLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 52.9 | 61.3 | 52.2 | 31.4 | |
| ALALoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 52.4 | 62.4 | 49.1 | 35.7 | |
| BS + CMOBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 52.3 | 62 | 49.1 | 36.7 | |
| LADELoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 51.9 | 62.3 | 49.3 | 31.2 | |
| Causal NormBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 51.8 | 62.7 | 48.8 | 31.6 | |
| De-confoundLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 51.8 | 62.7 | 48.8 | 31.6 | |
| LWSBackbone=ResNeXt-50, Stage 1=ACE, Strategy=Stage 2 Decoupled (LWS)2019.10 | 51.7 | 60.7 | 49.7 | 33.1 | |
| BSCUDA=true2023.02 | 51.6 | 61.9 | 49.2 | 32.3 | |
| cRTBackbone=ResNeXt-50, Stage 1=ACE, Strategy=Stage 2 Decoupled (cRT)2019.10 | 51.4 | 63.2 | 48.1 | 29.7 | |
| CE-DRW + CMOBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 51.4 | 60.8 | 48.6 | 35.5 | |
| LDAM-DRWCUDA=true2023.02 | 51.4 | 63.1 | 48 | 31.1 | |
| LDAM-DRW + CMOBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 51.1 | 62 | 47.4 | 30.8 | |
| Balanced Softmax (BS)Backbone=ResNet-50, Training Epochs=90 or 1002021.12 | 51 | 60.9 | 48.8 | 32.1 | |
| CE-DRWCUDA=true2023.02 | 51 | 61.8 | 48.3 | 30.3 | |
| BSCUDA=false2023.02 | 50.9 | 61.1 | 48.5 | 31.8 | |
| LDAM+DRWLoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 50.7 | 61.8 | 47.2 | 31.4 | |
| CRTCUDA=true2023.02 | 50.2 | 62.3 | 47.2 | 28.4 | |
| CE-DRWBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 50.1 | 61.7 | 47.3 | 28.8 | |
| CE-DRWCUDA=false2023.02 | 50.1 | 61.7 | 47.1 | 29 | |
| LWSBackbone=ResNeXt-50, Strategy=Stage 2 Decoupled (LWS)2019.10 | 49.9 | 60.2 | 47.2 | 30.3 | |
| LDAM-DRWBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 49.8 | 60.4 | 46.9 | 30.7 | |
| LDAM-DRWCUDA=false2023.02 | 49.8 | 60.4 | 46.9 | 30.7 | |
| cRTBackbone=ResNeXt-50, Strategy=Stage 2 Decoupled (cRT)2019.10 | 49.6 | 61.8 | 46.2 | 27.4 | |
| cRTLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 49.6 | 61.8 | 46.2 | 27.3 | |
| τ-normBackbone=ResNeXt-50, Strategy=Stage 2 Decoupled (τ-norm)2019.10 | 49.4 | 59.1 | 46.9 | 30.7 | |
| T-normLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 49.4 | 59.1 | 46.9 | 30.7 | |
| CE + CMOBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 49.1 | 67 | 42.3 | 20.5 | |
| cRTBackbone=ResNeXt-50, Stage 1=PC, Strategy=Stage 2 Decoupled (cRT)2019.10 | 48.9 | 59.3 | 46.1 | 29.5 | |
| NCMBackbone=ResNeXt-50, Stage 1=ACE, Strategy=Stage 2 Decoupled (NCM)2019.10 | 48.9 | 57.9 | 46.7 | 31 | |
| RemixBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 48.6 | 60.4 | 46.9 | 30.7 | |
| LWSBackbone=ResNeXt-50, Stage 1=PC, Strategy=Stage 2 Decoupled (LWS)2019.10 | 48.4 | 57.3 | 46.4 | 29.8 | |
| Decouple-LWSBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 47.7 | 57.1 | 45.2 | 29.3 | |
| LWSCUDA=false2023.02 | 47.7 | 57.1 | 45.2 | 29.3 | |
| ResNeXt-50 + ACEBackbone=ResNeXt-50, Strategy=Stage 1 with ACE2019.10 | 47.5 | 67.5 | 42.1 | 10.2 | |
| NCMBackbone=ResNeXt-50, Strategy=Stage 2 Decoupled (NCM)2019.10 | 47.3 | 56.6 | 45.3 | 28.1 | |
| Decouple-cRTBackbone=ResNet-50, Training Epochs=90 or 1002021.12 | 47.3 | 58.8 | 44 | 26.1 | |
| CECUDA=true2023.02 | 47.3 | 67.2 | 47 | 13.5 | |
| CRTCUDA=false2023.02 | 47.3 | 58.8 | 44 | 26.1 | |
| NCMBackbone=ResNeXt-50, Stage 1=PC, Strategy=Stage 2 Decoupled (NCM)2019.10 | 44.6 | 52.3 | 42.9 | 28.7 | |
| ResNeXt-50Backbone=ResNeXt-50, Strategy=Baseline Stage 12019.10 | 44.4 | 65.9 | 37.5 | 7.7 | |
| ResNeXt-50 + PCBackbone=ResNeXt-50, Strategy=Stage 1 with PC2019.10 | 42.8 | 63.9 | 35.5 | 8.8 | |
| CELoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 42.7 | 62.5 | 36.6 | 12.5 | |
| Cross Entropy (CE)Backbone=ResNet-50, Training Epochs=90 or 1002021.12 | 41.6 | 64 | 33.8 | 5.8 | |
| CECUDA=false2023.02 | 41.6 | 64 | 33.8 | 5.8 | |
| LFMRLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 37.2 | 47.1 | 35 | 17.5 | |
| OLTRLoss-ori.=false, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 35.6 | 43.2 | 35.1 | 18.5 | |
| CE+CBLoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 33.2 | 39.6 | 32.7 | 16.8 | |
| FocalLoss-ori.=true, Backbone=ResNet, Training Protocol=Training ResNets2023.10 | 30.5 | 36.4 | 29.9 | 16 |