Table 3.
Class-incremental action recognition performance on Kinetics-400 dataset. Mean-named columns show the average recognition accuracy (%) of all incremental learning steps. The other values report the recognition accuracy of one model.
| Exemplar | Base model | 20 × 5 stages | 10 × 10 stages | 5 × 20 stages | Static mode | ||||
|---|---|---|---|---|---|---|---|---|---|
| Methods | Mean | Final | Mean | Final | Mean | Final | |||
| 3Stream + cD | 15 | 73.5 | 64.42 | 60.26 | 63.02 | 58.34 | 60.74 | 55.23 | 69.84 |
| 3Stream + cD + tfD | 15 | 73.5 | 68.25 | 65.78 | 66.83 | 63.36 | 64.86 | 61.73 | 69.84 |
| 3Stream-D | 0 | 73.5 | 55.3 | 46.32 | 52.26 | 43.54 | 48.73 | 39.6 | 69.84 |
| 3Stream + cD | 0 | 73.5 | 60.4 | 56.86 | 58.7 | 54.21 | 55.38 | 51.63 | 69.84 |
| 3Stream + cD + tfD | 0 | 73.5 | 63.9 | 61.42 | 62.43 | 60.12 | 60.9 | 58.35 | 69.84 |