Missunderstanding of Torchvision FasterRCNN arguments


I just look at the FasterRCNN code on PyTorch repo and I think I misunderstood some arguments. FasterRCNN have box_nms_thresh as argument, but when testing different inference by setting up the nms with different values, I notice that it actually refer to an IOU threshold, because it will remove boxes that overlap more than the threshold passed.

In my understanding, I think nms threshold have to refer to the task of removing boxes lower than a confidence score (like the box_score_thresh do).

I have a doubt on this one, I would like someone confirm my understanding of the arguments here or explain me how should I understand it.

Thank you


I also look at the torchvision.ops implementation but here my thought is confirmed by passing iou_thresh as argument name torchvision.ops.nms( *boxes* , *scores* , *iou_threshold* ).

As per the source code, box_score_thresh (float): during inference, only return proposals with a classification score greater than box_score_thresh.
So it actually talks about classification scores. Not IoU scores.