Table 8 Ablation study of loss weight of \(L_{FMask}\) with RetinaNet-R50 as student, RetinaNet-X101 as teacher.
From: Instance mask alignment for object detection knowledge distillation
Method | mAP | \(AP_{50}\) | \(AP_{75}\) | \(AP_S\) | \(AP_M\) | \(AP_L\) |
|---|---|---|---|---|---|---|
Baseline | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
0.1 | 40.7 | 60.2 | 43.4 | 23.8 | 44.6 | 53.9 |
0.5 | 40.9 | 60.7 | 43.4 | 23.5 | 44.9 | 53.9 |
1 | 41.0 | 60.2 | 43.6 | 23.0 | 45.2 | 55.0 |
5 | 39.2 | 58.1 | 42.2 | 22.4 | 43.2 | 51.0 |
10 | 38.5 | 57.8 | 41.0 | 21.8 | 42.5 | 50.8 |
