Metrics

In terms of mAP "enclose" and "argmax" perform close to each other but "enclose" clearly beats "argmax" in precision

Key finding: if mode is not "argmax" i.e. NMS is baked into inference model, "num_detections" needs to be a lot larger in order to not drop any detections!

Now "enclose" is the obvious NMS mode to use with this type of evaluation.