Class-Incremental Learning on ImageNet (800 base, 40 novel, test)
57.82Accuracy (All Classes)score fusion (ours) best-Accall
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| score fusion (ours) best-AccallBackbone=ResNet102022.04 | 57.82 | 57.8 | 58.23 | 58.02 | |
| pre-trained modelBackbone=ResNet102022.04 | 57.69 | 58.03 | 51.2 | 54.62 | |
| score fusion (ours) best-balancedBackbone=ResNet102022.04 | 55.66 | 54.81 | 71.9 | 63.36 | |
| score fusion (ours) best-AccavgBackbone=ResNet102022.04 | 53.42 | 52.28 | 75.13 | 63.7 | |
| learning-based routing w/ Lrt-balBackbone=ResNet102022.04 | 52.82 | 52.12 | 66.1 | 59.11 | |
| joint learning (oracle)Backbone=ResNet102022.04 | 47.63 | 47.32 | 53.6 | 50.46 | |
| confidence-based routingBackbone=ResNet102022.04 | 35.43 | 33.01 | 81.5 | 57.25 | |
| RMBackbone=ResNet182022.04 | 21.89 | 28.78 | 80.9 | 49.84 | |
| RMBackbone=ResNet102022.04 | 21.1 | 18.12 | 77.7 | 47.91 | |
| iCaRLBackbone=ResNet102022.04 | 20.23 | 19.77 | 28.9 | 24.34 | |
| iCaRLBackbone=ResNet182022.04 | 18.12 | 17.04 | 38.6 | 27.82 | |
| fine-tuningBackbone=ResNet102022.04 | 3.86 | 0 | 77.2 | 38.6 |