Table 1 Main results on object detection. We use AP on different settings to evaluate results. Res101, Res50 represents using ResNet101 and ResNet50 as backbones.
From: Instance mask alignment for object detection knowledge distillation
| Â | Method | mAP | \(AP_{50}\) | \(AP_{75}\) | \(AP_S\) | \(AP_M\) | \(AP_L\) |
|---|---|---|---|---|---|---|---|
Teacher | FCOS-Res101 | 40.8 | 60.0 | 44.0 | 24.2 | 44.3 | 52.4 |
Student | FCOS-Res50 | 38.5 | 57.7 | 41.0 | 21.9 | 42.8 | 48.6 |
| Â | GID8 | 42.0 | 60.4 | 45.5 | 25.6 | 45.8 | 54.2 |
| Â | FRS9 | 40.9 | 60.3 | 43.6 | 25.7 | 45.2 | 51.2 |
|  | FGD10 | 42.1 | – | – | 27.0 | 46.0 | 54.6 |
| Â | IMA (Ours) | 42.4 | 61.0 | 45.8 | 26.6 | 45.9 | 54.8 |
Teacher | Faster RCNN-Res101 | 39.8 | 60.1 | 43.3 | 22.5 | 43.6 | 52.8 |
Student | Faster RCNN-Res50 | 38.4 | 59.0 | 42.0 | 21.5 | 42.1 | 50.3 |
| Â | KD-Zero11 | 38.4 | 59.4 | 41.7 | 22.7 | 41.8 | 45.9 |
| Â | FitNet12 | 38.8 | 59.6 | 41.8 | 22.3 | 42.2 | 50.7 |
| Â | FGFI13 | 39.4 | 60.3 | 43.0 | 22.9 | 42.5 | 52.0 |
|  | FGD10 | 40.4 | – | – | 22.8 | 44.5 | 53.5 |
| Â | IMA (Ours) | 40.6 | 60.9 | 43.9 | 23.0 | 44.5 | 54.0 |
Teacher | RetinaNet101-Res101 | 38.9 | 58.0 | 41.5 | 21.0 | 42.8 | 52.4 |
Student | RetinaNet50-Res50 | 37.4 | 56.7 | 39.6 | 20.0 | 40.7 | 49.7 |
| Â | KD-Zero11 | 36.8 | 56.6 | 39.4 | 21.9 | 40.6 | 48.2 |
| Â | FitNet12 | 36.3 | 56.0 | 39.0 | 20.1 | 40.3 | 47.1 |
| Â | FGFI13 | 37.3 | 57.1 | 40.0 | 21.0 | 41.5 | 49.7 |
|  | FGD10 | 39.6 | – | – | 22.9 | 44.3 | 53.4 |
| Â | IMA (Ours) | 39.7 | 58.6 | 41.4 | 22.7 | 42.9 | 51.3 |
