Long-Tail Image Classification on CIFAR100-LT imbalance ratio 100 (test)
80.3AccuracyCVS+ADRW
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| CVS+ADRWBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 80.3 | — | — | — | 79.7 | 80.5 | 80.7 | — | |
| CVS+ADRWBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 80 | — | — | — | 87.1 | 79.8 | 72 | — | |
| CVSBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 79.9 | — | — | — | 82.1 | 81.5 | 75.4 | — | |
| VS+TLA+ADRWBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 79.8 | — | — | — | 81.9 | 79.8 | 77.2 | — | |
| CE+ADRWBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 79.5 | — | — | — | 82 | 79.5 | 76.5 | — | |
| CVSBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 79.4 | — | — | — | 87.3 | 80.8 | 68.4 | — | |
| VS+TLA+ADRWBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 79.2 | — | — | — | 87.2 | 78.7 | 70.5 | — | |
| VSBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 79.1 | — | — | — | 88.4 | 79.3 | 67.9 | — | |
| LDAM+DRWBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 78.6 | — | — | — | 84.8 | 78.9 | 70.9 | — | |
| VSBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 77.7 | — | — | — | 82.1 | 80.4 | 69.4 | — | |
| CE+ADRWBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 73.4 | — | — | — | 88 | 76.7 | 52.6 | — | |
| LDAM+DRWBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 71.8 | — | — | — | 87.4 | 75.6 | 48.9 | — | |
| CEBackbone=ViT, Training Strategy=Full Fine-Tuning (FFT)2023.10 | 71.2 | — | — | — | 90.4 | 74.1 | 45.5 | — | |
| CEBackbone=ViT, Training Strategy=PEFT (LIFT)2023.10 | 66.6 | — | — | — | 87.9 | 70.7 | 36.9 | — | |
| CVS+ADRWBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 54.6 | — | — | — | 66.9 | 56.5 | 37.8 | — | |
| CVSBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 54.5 | — | — | — | 69.2 | 56.1 | 35.4 | — | |
| VSBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 54.1 | — | — | — | 68.4 | 56.6 | 34.5 | — | |
| VS+TLA+ADRWBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 53.2 | — | — | — | 66.7 | 56.3 | 37 | — | |
| CE+ADRWBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 53.1 | — | — | — | 68.3 | 56 | 31.9 | — | |
| LDAM+DRWBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 52.9 | — | — | — | 68.4 | 55.9 | 31 | — | |
| PaCo2023.03 | 51.24 | 45.29 | 36.42 | 5 | — | — | — | — | |
| PaCo + GMLEnsemble=false2023.03 | 50.53 | 45.47 | 39.2 | 9 | — | — | — | — | |
| PaCo + GMLEnsemble=true2023.03 | 49.82 | 45.7 | 41.02 | 15 | — | — | — | — | |
| RIDE2023.03 | 48.64 | 38.71 | 23.86 | 2 | — | — | — | — | |
| MiSLAS2023.03 | 47.05 | 40.16 | 30.93 | 5 | — | — | — | — | |
| CEBackbone=ResNet, Training Strategy=From scratch, Protocol=(c)2023.10 | 46.7 | — | — | — | 76.7 | 47 | 11.1 | — | |
| DIVE2023.03 | 45.11 | 37.08 | 25.18 | 2 | — | — | — | — | |
| LADE2023.03 | 44.39 | 39.05 | 32.58 | 5 | — | — | — | — | |
| LDAM2023.03 | 43.51 | 33.61 | 21.49 | 3 | — | — | — | — | |
| BSCE2023.03 | 42.24 | 35.16 | 26.21 | 3 | — | — | — | — | |
| CE + GML2023.03 | 41.06 | 36.59 | 31.26 | 6 | — | — | — | — | |
| CE2023.03 | 38.74 | 21.03 | 2.05 | 0 | — | — | — | — | |
| BBNMFlops=74.3 (1.1x)2020.10 | — | — | — | — | — | — | — | 42.6 | |
| Cross Entropy (CE) †MFlops=69.5 (1.0x), reproduced=true2020.10 | — | — | — | — | 66.1 | 37.3 | 10.6 | 39.1 | |
| Cross Entropy (CE) ‡MFlops=69.5 (1.0x), source=copied from (Cao et al., 2019)2020.10 | — | — | — | — | — | — | — | 38.3 | |
| cRT †MFlops=69.5 (1.0x), reproduced=true2020.10 | — | — | — | — | 64 | 44.8 | 18.1 | 43.3 | |
| Focal Loss ‡MFlops=69.5 (1.0x)2020.10 | — | — | — | — | — | — | — | 38.4 | |
| LDAM + DRWMFlops=69.5 (1.0x)2020.10 | — | — | — | — | — | — | — | 42 | |
| LDAM + DRW †MFlops=69.5 (1.0x), reproduced=true2020.10 | — | — | — | — | 61.5 | 41.7 | 20.2 | 42 | |
| LFME2020.10 | — | — | — | — | — | — | — | 43.8 | |
| M2m2020.10 | — | — | — | — | — | — | — | 43.5 | |
| OLTR †reproduced=true2020.10 | — | — | — | — | 61.8 | 41.4 | 17.6 | 41.2 | |
| RIDE (2 experts)MFlops=64.8 (0.9x), number of experts=22020.10 | — | — | — | — | 67.9 | 48.4 | 21.8 | 47 | |
| RIDE (3 experts)MFlops=77.8 (1.1x), number of experts=32020.10 | — | — | — | — | 68.1 | 49.2 | 23.9 | 48 | |
| RIDE (4 experts)MFlops=91.9 (1.3x), number of experts=42020.10 | — | — | — | — | 69.3 | 49.3 | 26 | 49.1 | |
| τ-norm †MFlops=69.5 (1.0x), reproduced=true2020.10 | — | — | — | — | 65.7 | 43.6 | 17.3 | 43.2 |