Anchors scaling for each feature map output

[rjtshrm]

pytorch-yolo-v3/util.py

Line 32 in fbb4ef9

anchors = [(a[0]/stride, a[1]/stride) for a in anchors]

I have a question since yolov3 has anchors for all three different scales. (as they mentioned in paper). Why again we need to down sample the anchors for each scale. It is kind of bit hard to understand the anchors scaling as I am new to anchor based detection.