Class-Incremental Learning on ImageNet disjoint-CIL 800+40
70.32Acc (All)joint learning (oracle)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| joint learning (oracle)Backbone=ResNet18, # of params=11.2M2022.04 | 70.32 | 70.43 | 68.2 | 69.32 | |
| score fusion (ours) best-AccallBackbone=ResNet18, # of params=19.6M2022.04 | 69.45 | 70.01 | 58.13 | 64.07 | |
| score fusion (ours, fc-only) best-AccallBackbone=ResNet18, # of params=11.2M2022.04 | 68.79 | 69.58 | 53.07 | 61.32 | |
| score fusion (ours) best-balancedBackbone=ResNet18, # of params=19.6M2022.04 | 67.36 | 66.61 | 82.37 | 74.49 | |
| score fusion (ours, fc-only) best-balancedBackbone=ResNet18, # of params=11.2M2022.04 | 66.76 | 66.5 | 71.83 | 69.17 | |
| score fusion (ours, fc-only) best-AccavgBackbone=ResNet18, # of params=11.2M2022.04 | 65.89 | 65.49 | 73.77 | 69.63 | |
| score fusion (ours) best-AccavgBackbone=ResNet18, # of params=19.6M2022.04 | 65.83 | 64.85 | 82.5 | 75.17 | |
| joint learning (oracle)Backbone=ResNet10, # of params=4.9M2022.04 | 63.8 | 63.94 | 61 | 62.47 | |
| score fusion (ours) best-AccallBackbone=ResNet10, # of params=8.6M2022.04 | 63.24 | 63.77 | 52.67 | 58.22 | |
| score fusion (ours, fc-only) best-AccallBackbone=ResNet10, # of params=4.9M2022.04 | 62.65 | 63.56 | 44.53 | 54.05 | |
| score fusion (ours) best-balancedBackbone=ResNet10, # of params=8.6M2022.04 | 62.15 | 61.49 | 75.37 | 68.43 | |
| score fusion (ours, fc-only) best-balancedBackbone=ResNet10, # of params=4.9M2022.04 | 61.01 | 60.81 | 65.07 | 62.94 | |
| score fusion (ours) best-AccavgBackbone=ResNet10, # of params=8.6M2022.04 | 58.9 | 57.73 | 82.4 | 70.06 | |
| score fusion (ours, fc-only) best-AccavgBackbone=ResNet10, # of params=4.9M2022.04 | 57.91 | 57.24 | 71.57 | 64.4 | |
| WABackbone=ResNet182022.04 | 54.79 | 55.17 | 47.2 | 51.19 | |
| DER w/o PBackbone=ResNet10, # of params=9.8M2022.04 | 52.31 | 52.43 | 50.1 | 51.27 | |
| WABackbone=ResNet102022.04 | 51.33 | 52.33 | 31.4 | 41.87 | |
| BiCBackbone=ResNet182022.04 | 31.5 | 28.75 | 86.6 | 57.68 | |
| BiCBackbone=ResNet102022.04 | 30.3 | 27.55 | 85.2 | 56.38 | |
| iCaRLBackbone=ResNet102022.04 | 16.26 | 13.91 | 63.4 | 38.66 | |
| iCaRLBackbone=ResNet182022.04 | 10.65 | 8.15 | 60.8 | 34.78 | |
| LwFBackbone=ResNet102022.04 | 9.5 | 5.54 | 88.53 | 47.04 | |
| LwFBackbone=ResNet182022.04 | 9.5 | 5.46 | 90.3 | 47.88 | |
| fine-tuningBackbone=ResNet18, # of params=11.2M2022.04 | 4.25 | 0 | 89.37 | 44.68 | |
| fine-tuningBackbone=ResNet10, # of params=4.9M2022.04 | 4.18 | 0.01 | 87.63 | 43.82 |