Image Classification on ImageNet LT
82Top-1 AccuracyCategory Extrapolation (Ours)
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Category Extrapolation (Ours)Pre-training paradigm=DINOv2, Data configuration=proposed method2024.10 | 82 | — | 84.7 | 81.5 | 76.2 | — | — | — | — | — | |
| Baseline + SDPre-training paradigm=DINOv2, Data configuration=selected neighbor categories2024.10 | 80.5 | — | 83.8 | 79.8 | 73.4 | — | — | — | — | — | |
| BaselinePre-training paradigm=DINOv2, Data configuration=standard2024.10 | 79.6 | — | 84.3 | 78.3 | 71.1 | — | — | — | — | — | |
| Category Extrapolation (Ours)Pre-training paradigm=CLIP, Data configuration=proposed method2024.10 | 77.3 | — | 79.1 | 76.8 | 74.1 | — | — | — | — | — | |
| Baseline + RDPre-training paradigm=DINOv2, Data configuration=random auxiliary data2024.10 | 77.2 | — | 83.3 | 75.7 | 65.4 | — | — | — | — | — | |
| Baseline + SDPre-training paradigm=CLIP, Data configuration=selected neighbor categories2024.10 | 75.2 | — | 77.8 | 74.2 | 71.3 | — | — | — | — | — | |
| BaselinePre-training paradigm=CLIP, Data configuration=standard2024.10 | 74 | — | 77.2 | 72.8 | 68.5 | — | — | — | — | — | |
| Baseline + RDPre-training paradigm=CLIP, Data configuration=random auxiliary data2024.10 | 68.8 | — | 75.4 | 67.4 | 55.2 | — | — | — | — | — | |
| Category Extrapolation (Ours)Pre-training paradigm=Scratch, Data configuration=proposed method2024.10 | 68.2 | — | 74.5 | 66.2 | 57.4 | — | — | — | — | — | |
| Baseline + SDPre-training paradigm=Scratch, Data configuration=selected neighbor categories2024.10 | 64.9 | — | 73.4 | 62.1 | 50.6 | — | — | — | — | — | |
| BaselinePre-training paradigm=Scratch, Data configuration=standard2024.10 | 60.9 | — | 72.9 | 56.8 | 41.4 | — | — | — | — | — | |
| ConCutMixBackbone=ResNext-50, Training Epochs=1802024.07 | 60.3 | — | 71.5 | 57.2 | 40 | — | — | — | — | — | |
| 3LSSLBackbone=ResNeXt-502023.02 | 59.9 | — | 70 | 57.8 | 38.7 | — | — | — | — | — | |
| ConCutMixBackbone=ResNext-50, Training Epochs=902024.07 | 59.7 | — | 70.7 | 56.6 | 39.8 | — | — | — | — | — | |
| 3LSSLBackbone=ResNet-502023.02 | 59.1 | — | 68.5 | 57.6 | 38.3 | — | — | — | — | — | |
| OursBackbone=ResNeXt-502026.03 | 58.9 | — | 69.7 | 56 | 38.6 | — | — | — | — | — | |
| SADEBackbone=ResNeXt-502023.02 | 58.8 | — | 66.5 | 57 | 43.5 | — | — | — | — | — | |
| GBGBackbone=ResNeXt-502026.03 | 58.7 | — | 69.6 | 55.8 | 38.1 | — | — | — | — | — | |
| BCL + CutMixBackbone=ResNext-50, Training Epochs=1802024.07 | 58.3 | — | 67.7 | 56.2 | 39 | — | — | — | — | — | |
| PaCoBackbone=ResNeXt-502023.02 | 58.2 | — | 67.5 | 56.9 | 36.7 | — | — | — | — | — | |
| ProCoBackbone=ResNeXt-50, Epochs=902024.03 | 58 | — | — | — | — | — | — | — | — | — | |
| BCL + CutMixBackbone=ResNext-50, Training Epochs=902024.07 | 57.9 | — | 68.5 | 54.9 | 38.8 | — | — | — | — | — | |
| RIDE (4*) + CRBackbone=ResNext-502023.03 | 57.8 | — | — | — | — | — | — | 68.5 | 54.2 | 38.8 | |
| MBJ + RIDE (4 experts)Backbone=ResNeXt-502020.08 | 57.7 | — | 68.4 | 54.1 | 37.7 | — | — | — | — | — | |
| FeatReconBackbone=ResNeXt-502026.03 | 57.5 | — | 67.9 | 54.7 | 37.8 | — | — | — | — | — | |
| RIDE + CMO + CRBackbone=ResNext-502023.03 | 57.4 | — | — | — | — | — | — | 67.3 | 54.6 | 38.4 | |
| ProCoBackbone=ResNet-50, Epochs=902024.03 | 57.3 | — | — | — | — | — | — | — | — | — | |
| BCL + MixupBackbone=ResNext-50, Training Epochs=1802024.07 | 57.3 | — | 67.6 | 54.6 | 38.3 | — | — | — | — | — | |
| GLMC + BSBackbone=ResNeXt-502023.05 | 57.21 | — | 64.76 | 55.67 | 42.19 | — | — | — | — | — | |
| BCL + MixupBackbone=ResNext-50, Training Epochs=902024.07 | 57.2 | — | 67.6 | 53.9 | 39.9 | — | — | — | — | — | |
| BCLBackbone=ResNext-50, Training Epochs=1802024.07 | 57.2 | — | 66.9 | 54.7 | 38.6 | — | — | — | — | — | |
| BCL + CMOBackbone=ResNext-50, Training Epochs=1802024.07 | 57.2 | — | 65.3 | 55.9 | 39.2 | — | — | — | — | — | |
| BCLBackbone=ResNeXt-502023.02 | 57.1 | — | 67.9 | 54.2 | 36.6 | — | — | — | — | — | |
| BCLBackbone=ResNeXt-50, Training Epochs=1802023.05 | 57.1 | — | 67.9 | 54.2 | 36.6 | — | — | — | — | — | |
| BCL + CMOBackbone=ResNext-50, Training Epochs=902024.07 | 57.1 | — | 66.9 | 54.7 | 39.3 | — | — | — | — | — | |
| PaCoBackbone=ResNet-502023.02 | 57 | — | 65 | 53.7 | 38.2 | — | — | — | — | — | |
| RIDE (4 experts)Backbone=ResNeXt-502023.02 | 56.8 | — | 68.2 | 53.8 | 36 | — | — | — | — | — | |
| Baseline + RDPre-training paradigm=Scratch, Data configuration=random auxiliary data2024.10 | 56.8 | — | 72.1 | 50.4 | 35.8 | — | — | — | — | — | |
| RIDE (4 experts)Backbone=ResNeXt-50, GFlops=5.19 (1.2x), number of experts=42020.10 | 56.8 | — | 68.2 | 53.8 | 36 | — | — | — | — | — | |
| BCLBackbone=ResNeXt-50, Training Epochs=902023.05 | 56.7 | — | 67.2 | 53.9 | 36.5 | — | — | — | — | — | |
| GLMC + MaxNormBackbone=ResNeXt-502023.05 | 56.7 | — | 60.8 | 55.9 | 45.5 | — | — | — | — | — | |
| BCLBackbone=ResNeXt-50, Epochs=902024.03 | 56.7 | — | — | — | — | — | — | — | — | — | |
| BCLBackbone=ResNext-50, Training Epochs=902024.07 | 56.7 | — | 67.2 | 53.9 | 36.5 | — | — | — | — | — | |
| BCLBackbone=ResNeXt-502026.03 | 56.7 | — | — | — | — | — | — | — | — | — | |
| RIDE (4 experts)Backbone=ResNeXt-502020.08 | 56.6 | — | 67.8 | 53.4 | 36.2 | — | — | — | — | — | |
| RIDE (4*)Backbone=ResNext-502023.03 | 56.6 | — | — | — | — | — | — | 67.8 | 53.4 | 36.2 | |
| Logit Adj.Backbone=ResNeXt-50, Epochs=902024.03 | 56.5 | — | — | — | — | — | — | — | — | — | |
| RIDEBackbone=ResNeXt-502021.10 | 56.4 | — | — | — | — | — | — | — | — | — | |
| RIDEBackbone=ResNeXt-50, Epochs=902024.03 | 56.4 | — | — | — | — | — | — | — | — | — | |
| RIDE (3 experts)Backbone=ResNeXt-50, GFlops=4.69 (1.1x), number of experts=32020.10 | 56.4 | — | 67.6 | 53.5 | 35.9 | — | — | — | — | — | |
| GLMCBackbone=ResNeXt-502023.05 | 56.3 | — | 70.1 | 52.4 | 30.4 | — | — | — | — | — | |
| Swin-S + RIDE (2 experts)GFlops=8.32 (1.0x)2020.10 | 56.3 | — | 67.4 | 52.9 | 37 | — | — | — | — | — | |
| RIDE (3 experts)+CMOBackbone=ResNet-502023.02 | 56.2 | — | 66.4 | 53.9 | 35.6 | — | — | — | — | — | |
| RIDE (3 experts) + CMOBackbone=ResNet-502023.05 | 56.2 | — | 66.4 | 53.9 | 35.6 | — | — | — | — | — | |
| RIDE + CMOBackbone=ResNext-502023.03 | 56.2 | — | — | — | — | — | — | 66.1 | 54.9 | 35.8 | |
| ResLTBackbone=ResNeXt-50, Epochs=902024.03 | 56.1 | — | — | — | — | — | — | — | — | — | |
| SSDBackbone=ResNeXt-502023.02 | 56 | — | 66.8 | 53.1 | 35.4 | — | — | — | — | — | |
| PaCoBackbone=ResNeXt-50, Training Epochs=1802023.05 | 56 | — | 64.4 | 55.7 | 33.7 | — | — | — | — | — | |
| SSDBackbone=ResNeXt-502023.05 | 56 | — | 66.8 | 53.1 | 35.4 | — | — | — | — | — | |
| RIDE (3 experts)Backbone=Swin-S2023.05 | 56 | — | 66.9 | 52.8 | 37.4 | — | — | — | — | — | |
| BCLBackbone=ResNet-50, Epochs=902024.03 | 56 | — | — | — | — | — | — | — | — | — | |
| PaCoBackbone=ResNext-50, Training Epochs=1802024.07 | 56 | — | 64.4 | 55.7 | 33.7 | — | — | — | — | — | |
| Swin-S + RIDE (3 experts)GFlops=9.68 (1.1x)2020.10 | 56 | — | 66.9 | 52.8 | 37.4 | — | — | — | — | — | |
| RIDE (2 experts)Backbone=ResNeXt-50, GFlops=3.92 (0.9x), number of experts=22020.10 | 55.9 | — | 67.6 | 52.5 | 35 | — | — | — | — | — | |
| RIDE (2 experts)Backbone=ResNeXt-502026.03 | 55.9 | — | — | — | — | — | — | — | — | — | |
| MoCo V2+rwSAMBackbone=ResNet-502021.10 | 55.5 | — | — | — | — | — | — | — | — | — | |
| RIDE (4 experts)Backbone=ResNet-502023.02 | 55.4 | — | 66.2 | 52.3 | 36.5 | — | — | — | — | — | |
| Balanced SoftmaxBackbone=ResNeXt-50, Training Epochs=1802023.05 | 55.4 | — | 65.8 | 53.2 | 34.1 | — | — | — | — | — | |
| BALMSBackbone=ResNext-50, Training Epochs=1802024.07 | 55.4 | — | 65.8 | 53.2 | 34.1 | — | — | — | — | — | |
| ResLTBackbone Model=ResNeXt-101-32x4d, One-stage=true2021.01 | 55.1 | — | 63.3 | 53.3 | 40.3 | — | — | — | — | — | |
| Logit Adj.Backbone=ResNet-50, Epochs=902024.03 | 55.1 | — | — | — | — | — | — | — | — | — | |
| MoCo V2Backbone=ResNet-502021.10 | 55 | — | — | — | — | — | — | — | — | — | |
| RIDEBackbone=ResNet-502021.10 | 54.9 | — | — | — | — | — | — | — | — | — | |
| GCLBackbone=ResNeXt-502023.02 | 54.9 | — | — | — | — | — | — | — | — | — | |
| GCLBackbone=ResNet-50, Epochs=902024.03 | 54.9 | — | — | — | — | — | — | — | — | — | |
| RIDEBackbone=ResNet-50, Epochs=902024.03 | 54.9 | — | — | — | — | — | — | — | — | — | |
| GCLBackbone=ResNext-502023.03 | 54.9 | — | — | — | — | — | — | — | — | — | |
| PaCoBackbone=ResNext-502023.03 | 54.4 | — | — | — | — | — | — | 63.2 | 51.6 | 39.2 | |
| Swin-T + RIDE (3 experts)GFlops=4.95 (1.1x)2020.10 | 54.2 | — | 65.5 | 50.5 | 35.1 | — | — | — | — | — | |
| LWSBackbone Model=ResNeXt-101-32x4d, One-stage=false2021.01 | 54 | — | 65.7 | 51.4 | 34.7 | — | — | — | — | — | |
| Difficulty-Net (Ours) + LASBackbone=ResNet-50, Training Protocol=decoupled learning2022.09 | 54 | — | — | — | — | — | — | — | — | — | |
| ICCLBackbone=ResNeXt-50, Epochs=902024.03 | 54 | — | — | — | — | — | — | — | — | — | |
| tau-normalizeBackbone Model=ResNeXt-101-32x4d, One-stage=false2021.01 | 53.9 | — | 65.3 | 51.5 | 35.2 | — | — | — | — | — | |
| weight balancing + MaxNormBackbone=ResNeXt-502023.05 | 53.9 | — | 62.5 | 50.4 | 41.5 | — | — | — | — | — | |
| SSDBackbone=ResNeXt-50, Epochs=902024.03 | 53.8 | — | — | — | — | — | — | — | — | — | |
| Difficulty-Net (Ours) + Bal. SoftmaxBackbone=ResNet-50, Training Protocol=e2e training2022.09 | 53.7 | — | — | — | — | — | — | — | — | — | |
| Difficulty-Net (Ours) + LWSBackbone=ResNet-50, Training Protocol=decoupled learning2022.09 | 53.7 | — | — | — | — | — | — | — | — | — | |
| vMF ClassifierBackbone=ResNeXt-50, Epochs=902024.03 | 53.7 | — | — | — | — | — | — | — | — | — | |
| Swin-T + RIDE (2 experts)GFlops=3.48 (0.8x)2020.10 | 53.6 | — | 65 | 50 | 34.1 | — | — | — | — | — | |
| PaCo + Bal. SoftmaxBackbone=ResNet-50, Training Protocol=e2e training, Reproduced=true2022.09 | 53.5 | — | — | — | — | — | — | — | — | — | |
| DRO-LTBackbone=ResNet-50, Training Protocol=decoupled learning2022.09 | 53.5 | — | — | — | — | — | — | — | — | — | |
| Difficulty-Net (Ours) + cRTBackbone=ResNet-50, Training Protocol=decoupled learning2022.09 | 53.5 | — | — | — | — | — | — | — | — | — | |
| DRO-LTBackbone=ResNet-502023.02 | 53.5 | — | 64 | 49.8 | 33.1 | — | — | — | — | — | |
| DisAlignBackbone=ResNeXt-50, Epochs=902024.03 | 53.4 | — | — | — | — | — | — | — | — | — | |
| DisalignBackbone=ResNext-50, Training Epochs=902024.07 | 53.4 | — | 62.7 | 52.1 | 31.4 | — | — | — | — | — | |
| MISLASBackbone=ResNext-502023.03 | 53.4 | — | — | — | — | — | — | 65.3 | 50.6 | 33 | |
| MBJ + CRBackbone=ResNext-502023.03 | 53.4 | — | — | — | — | — | — | 62.8 | 49.2 | 40.4 | |
| DisAlignBackbone=ResNeXt-502026.03 | 53.4 | — | 62.7 | 52.1 | 31.4 | — | — | — | — | — | |
| cRTBackbone Model=ResNeXt-101-32x4d, One-stage=false2021.01 | 53.3 | — | 66.2 | 50.4 | 30.8 | — | — | — | — | — | |
| DiVEBackbone=ResNeXt-502023.02 | 53.1 | — | 64.1 | 50.4 | 31.5 | — | — | — | — | — |