Table 5 Ablation studies on our proposed IMA using ResNet50-based RetinaNet with ResNeXt101-based RetinaNeXt serving as the teacher.
From: Instance mask alignment for object detection knowledge distillation
| Â | Method | mAP | \(AP_{50}\) | \(AP_{75}\) | \(AP_{S}\) | \(AP_{M}\) | \(AP_{L}\) |
|---|---|---|---|---|---|---|---|
Teacher | RetinaNet-ResNeXt101 | 41.6 | 61.4 | 44.3 | 23.9 | 45.5 | 54.5 |
Student | RetinaNet-Res50 | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| Â | + KD | 40.2 | 59.5 | 43.0 | 22.6 | 44.3 | 53.4 |
| Â | + Instance Mask | 40.7 | 60.2 | 43.4 | 23.8 | 44.6 | 53.9 |
| Â | + Standardization | 40.9 | 60.7 | 43.4 | 23.5 | 44.9 | 53.9 |
| Â | + Adaptive Scale | 41.0 | 60.2 | 43.6 | 23.0 | 45.2 | 55.0 |
