锚框:以每一个像素点为中心生成多个大小和宽高比不同的边界框
图像 高h , 宽w
大小为
锚框的高 h1 = hs/√r
锚框的宽 w1 = ws√r
s1,...,sn 与 r1,... ,rm 这样一张图的锚框有 whnm 个
但对于一点来说,我们感兴趣的框有:(s1,r1),(s1,r2)......(s1,rm),(s2,r1),......(sn,r1)
假如 w=416 h=416 s=[0.75,0.5,0.25] r=[1,2,,0.5] 在 (200,200 )处生成5个框:
[44, 44, 356, 356], [96, 96, 304, 304], [148, 148, 252, 252], [0, 90, 416, 310], [90, 0, 310, 416]
accard系数(Jaccard index)可以衡量两个集合的相似度。给定集合,它们的Jaccard系数即二者交集大小除以二者并集大小:
import math import numpy as np w = 416 h = 416 s = [0.75, 0.5, 0.25] r = [1, 2, 0.5] def create_box(x, y,imgw, imgh, sizes, ratios): # nn = int(len(sizes) + len(ratios) - 1) size_1 = sizes[0] ratio_1 = ratios[0] ratio_2 = ratios[1:] x = x / imgw y = y / imgh list_1 = [[ x - i * math.sqrt(ratio_1) / 2, y - i / (2 * math.sqrt(ratio_1)), x + i * math.sqrt(ratio_1) / 2, y + i / (2 * math.sqrt(ratio_1)) ] for i in sizes] list_2 = [[ x - size_1 * math.sqrt(i) / 2, y - size_1 / (2 * math.sqrt(i)), x + size_1 * math.sqrt(i) / 2, y + size_1 / (2 * math.sqrt(i)) ] for i in ratio_2] list_1.extend(list_2) list_all = [] for l in list_1: if l[0] < 0: l[0] = 0 else: l[0] = round(l[0] * imgw) if l[1] < 0: l[1] = 0 else: l[1] = round(l[1] * imgh) if l[2] > 1: l[2] = imgw else: l[2] = round(l[2] * imgw) if l[3] > 1: l[3] = imgh else: l[3] = round(l[3] * imgh) list_all.append(l) return list_all if __name__ == '__main__': print(create_box(200, 200, w, h, s, r))
本站文章如无特殊说明,均为本站原创,如若转载,请注明出处:目标检测——锚框——-anchor box - Python技术站