About the NMS

[DrugRui]

Thanks for your clean and useful code implementation.I would like to consult you why 'class_bboxes' on the NMS are not arranged in descending order of 'class_probs' in 'model.py' when validating or testing. The code is as follows: kept_indices = nms(class_bboxes, class_probs, threshold)

[potterhsu]

1. First of all, the function nms has handled boxes sorting
2. So that in generate_detections boxes doesn't require to sort before nms is called
3. You might be confused by similar procedure in generate_proposals which has done sorting before nms, it's because that we want to take top N boxes and then pass to nms for the efficiency (note that #boxes between generate_detections and generate_proposals was different, typically <1k and >10k)

[DrugRui]

Thank you for your reply. It is useful to me

[DrugRui]

I have another question: why is the threshold of detection_probs for the verification phase very low, so that there are enough boxes involved in the calculation of map? As in the code in evaluator.py: kept_indices = (detection_probs > 0.05).nonzero().view(-1)

[potterhsu]

Right, and this threshold is determined by reference to other popular repos, such as maskrcnn-benchmark

[DrugRui]

Well, thanks a lot for your help.