Table 1.
Performance comparison with existing methods Shown is the accuracy of various annotation methods on both datasets.
| Model | Accuracy (%) |
|---|---|
| Home-cage | |
| Human | 71.6 |
| CleverSys commercial system7 | 61.0 |
| Jhuang et al.7 | 78.3 |
| Le and Murari23 | 73.5 |
| Jiang et al.24 | 81.5 |
| DeepAction | 79.5 |
| CRIM13 | |
| Human | 69.7 |
| Burgos-Artizzu et al.6 | 62.6 |
| Eyjolfsdottir et al.25 | 37.2 |
| Zhang et al.26 | 61.9 |
| Meng et al.27 | 68.6 |
| DeepAction | 73.9% |
“Human” denotes the agreement between two human annotator groups (see “Methods: Inter-observer reliability” section). The accuracy for DeepAction on the home-cage and CRIM13 datasets is the mean accuracy from 12-fold and two-fold cross-validation, respectively, to provide a comparable reference to Jhuang et al.7 and Burgos-Artizzu et al.6 (see “Methods: Comparison with existing methods” section).