Metrics

All tests used NMS threshold 0.25, which seems to work well

"argmax" is clearly best for mAP and recall, for best precision one should use

• max_detections=1000
• top-k=50
• nms-mode="median"

There's a whopping +24% difference in precision with these settings vs. "argmax" at the expense of -4% recall

Though mAP metric seems to overly favor recall here... Although for SAR context recall is more important

Although median NMS seems to be very sensitive to the choice of num_detections and top-k, where "score_vote" is not, but gives slightly worse overall performance