Table 3.
Results of the ablation study for PoseNet on the AFLW and PoseTrack datasets.
| Model | AFLW dataset | PoseTrack dataset | ||||||
|---|---|---|---|---|---|---|---|---|
| Accuracy | Recall | F1 score | AUC | Accuracy | Recall | F1 score | AUC | |
| w./o. Graph-based temporal learning | 91.23 ± 0.02 | 89.45 ± 0.03 | 88.30 ± 0.02 | 90.01 ± 0.03 | 92.02 ± 0.03 | 89.78 ± 0.02 | 88.91 ± 0.02 | 89.67 ± 0.03 |
| w./o. Multi-modal data fusion | 92.15 ± 0.03 | 90.33 ± 0.02 | 89.45 ± 0.02 | 91.12 ± 0.02 | 93.11 ± 0.02 | 91.02 ± 0.03 | 90.34 ± 0.03 | 90.91 ± 0.02 |
| w./o. Domain-constrained optimization | 93.02 ± 0.02 | 91.22 ± 0.03 | 90.15 ± 0.02 | 91.90 ± 0.02 | 93.89 ± 0.03 | 91.76 ± 0.02 | 90.88 ± 0.03 | 91.22 ± 0.02 |
| Ours | 94.56 ± 0.02 | 92.34 ± 0.02 | 91.89 ± 0.03 | 92.14 ± 0.03 | 94.10 ± 0.03 | 92.87 ± 0.02 | 91.45 ± 0.03 | 91.72 ± 0.02 |
The values in bold are the best values.