Table 2 Detection results on the Taiping Houkui Tea dataset using different distillation methods and detection frameworks. Defeat can only be applied to anchor-based detectors.
From: Learning lightweight tea detector with reconstructed feature and dual distillation
Teacher | Student | AP (%) |
|---|---|---|
RetinaNet-Res101 (AP = 75.98%) | RetinaNet-Res50 | 73.23 |
Defeat18 | 75.66 | |
MGD20 | 75.92 | |
DiffKD53 | 76.25 | |
RFDD | 76.37(+ 3.14) | |
Faster RCNN-Res101 (AP = 77.46%) | Faster RCNN-Res50 | 74.15 |
Defeat18 | 76.96 | |
MGD20 | 77.31 | |
DiffKD53 | 77.59 | |
RFDD | 77.68(+ 3.53) | |
Mask RCNN-Res101 (AP = 80.12%) | Mask RCNN-Res50 | 77.92 |
Defeat18 | 79.35 | |
MGD20 | 79.67 | |
DiffKD53 | 79.88 | |
RFDD | 79.90(+ 1.98) | |
FCOS-Res101 (AP = 76.59%) | FCOS-Res50 | 73.81 |
MGD20 | 76.52 | |
DiffKD53 | 76.63 | |
RFDD | 76.76(+ 2.95) | |
RepPoints-Res101 (AP = 77.40%) | RepPoints-Res50 | 73.27 |
MGD20 | 77.24 | |
DiffKD53 | 77.46 | |
RFDD | 77.53(+ 3.26) |
