hi, guys~ I’m training my own dataset with mrcnn. But when I tested my model, I found some bboxes overlapped with each other like the picture below. I’m really confused about the principle here, shouldn’t the nms have suppressed these boxes after rpn before classification? Thanks a lot for any help!
Those bboxes are of the same class?
No, they are from different classes, and of course, some of them are bad predictions with relative low scores. 
BTY, I’m not sure, are there two nms during reference stage, one for all proposals in rpn, and the other only between the same class after classification or what?