1.2Faster RCNN源码解析(pytorch)

根据弹幕老哥提示

anchor

scale * scale = hs * ws =

那么可以根据 scale * scale = ws * ws * aspect_ratios 可得 ws = scale / 根号下 aspect_ratios 也就是说w_ratio= 1/根号下aspect_ratios

(scale / 根号下aspect_ratios )[也就是ws] * hs = scale *scale

可得 hs = scale * 根号下aspect_ratios

即h_ratios = 根号下aspect_ratios