Open WangLuran opened 1 year ago
Thanks for this contribution @WangLuran ! There seems to be a few conflicts that are preventing me from merging. Would you also please comment which functions have been edited (and how)?
OK, got it! I will merge the commits and write the comments
This PR has become stale. @PatBall1 is it still relevant?
Changing the nms algorithm in rpn layer by considering inter over area instead of inter over union. It inceases the speed of divergence and mask quality and a little in ap50. The overlapping of crowns of different types is also solved. Change in proprocessing in predict.py. Instead of just deleting all crowns with a objectness value less than a threshold, a changed nms algorithm is applied - we use inter over area of masks instead of bounding boxes for a better output.