Table 2. Experiment setup.
Batch Size | Image Size | Learning Rate | Weight Decay | NMS Threshold | Confidence Threshold | Epoch/Iter (2-class det) | Epoch/Iter (1-class det) | |
---|---|---|---|---|---|---|---|---|
Faster RCNN | 8 | 600 | 10−3 | 10−1 | 0.45 | 0.50 | 7 | 3 |
SSD | 8 | 300 | 4 × 10−4 | 10−4 | 0.45 | 0.50 | 8 | 35k iter |
YOLOv3 | 32 | 416 | 10−3 | 5 × 10−4 | 0.45 | 0.50 | 24k iter | |
RetinaNet | 1 | 608 | 10−5 | 0.50 | 0.50 | 9 | 1 | |
DetNet | 8 | 600 | 10−3 | 10−4 | 0.45 | 0.50 | 2 | 5 |
RefineDet | 8 | 512 | 10−4 | 5 × 10−4 | 0.45 | 0.50 | 35k iter | 130k iter |
YOLOv4 | 4 | 416 | 10−4 | 5 × 10−4 | 0.45 | 0.50 | 40 | 26 |
ATSS | 16 | 600 | 5 × 10−3 | 10−4 | 0.45 | 0.50 | 15k iter | 10k iter |