Det为预测框 Gt为真实框
import numpy as np
from functools import reduce
def iou_cal(boxa, boxb): #计算IOU
x1, y1, x2, y2 = boxa #a框的左上角和右下角坐标
x3, y3, x4, y4 = boxb #b框的左上角和右下角坐标
left_max = max(x1, x3) #计算交集框的左边的横坐标最大值
right_min = min(x2, x4) #计算交集框的右边的横坐标最小值
top_max = max(y1, y3) #计算交集框的上边的纵坐标最大值
bottom_min = min(y2, y4) #计算交集框下边的纵坐标最小值
#计算交集框的面积 长是左右横坐标的差值,宽是上下纵坐标的差值 当ab框没有交集时,左右横坐标或上下纵坐标差值为负值,此时取0即可
inter_area = max(0, right_min-left_max) * max(0, bottom_min-top_max)
#计算a框和b框的面积
a_area = (x2 - x1) * (y2 - y1)
b_area = (x4 - x3) * (y4 - y3)
#计算交并比
iou = inter_area / (a_area + b_area - inter_area)
return iou
cls_num = 3 #目标类别数量
pre_boxes = [] #保存预测框的list
gt_boxes = [] #保存真实框的list
cls_pre_boxes = {i: [] for i in range(cls_num)} #保存不同类别预测框的字典
cls_gt_boxes = {j: [] for j in range(cls_num)} #保存不同类别真实框的字典
#随机生成预测值和标签值 (标签和预测值一致)
for i in range(50):
x1 = np.random.randint(10, 50)
x2 = x1 + np.random.randint(20, 50)
y1 = np.random.randint(20, 60)
y2 = y1 + np.random.randint(30, 80)
cls = np.random.randint(0, 3)
conf = np.random.rand()
pre_boxes.append([x1, y1, x2, y2, cls, conf])
gt_boxes.append([x1, y1, x2, y2, cls, 0])
#将预测框和标签框按类别分类
for pre_box in pre_boxes:
cls_pre_boxes[pre_box[4]].append(pre_box)
for gt_box in gt_boxes:
cls_gt_boxes[gt_box[4]].append(gt_box)
res = [] #保存每个类别的ap
#按类别遍历全部预测框计算ap
for pre_cls, pre_boxes in cls_pre_boxes.items():
gt_cls = pre_cls #当前类别序号
gt_boxes = cls_gt_boxes[gt_cls] #当前类别的真实框
#对当前类别的预测值按置信度降序排序
pre_boxes = sorted(pre_boxes, key=lambda x: x[5], reverse=True)
TP = np.zeros(len(pre_boxes))
FP = np.zeros(len(pre_boxes))#数组保存每个预测框是TP还是FP
for ind, pre_box in enumerate(pre_boxes):
iou_max = 0 # 初始化当前预测框与真实框的最大IOU为0
for indx, gt_box in enumerate(gt_boxes):
iou = IOU_calculate.iou_cal(pre_box[:4], gt_box[:4])
if iou > iou_max:
iou_max = iou
jmax = indx #保存最大iou的真实框的序号
# 若当前预测框与真实框的最大iou大于0.5且匹配的真实框没有被匹配过则记为TP且标记这个真实框被匹配过
if iou_max > 0.5 and gt_boxes[jmax][5] == 0:
TP[ind] = 1
gt_boxes[jmax][5] = 1
else: #否则标记为FP
FP[ind] = 1
#计算累计的TP和FP 即有n个预测框生成对应的n个precision和recall以绘制PR曲线求ap
acc_TP = np.cumsum(TP)
acc_FP = np.cumsum(FP)
#计算截至到每个预测框的recall和precision
rec = acc_TP / len(gt_boxes)
acc = np.divide(acc_TP, acc_TP + acc_FP)
#计算ap
ap = np.trapz(acc, rec) #求PR曲线下面积 y轴acc, x轴rec
res.append((pre_cls, ap)) #将当前类别的ap保存
print(res)
mAP = reduce(lambda x, y: x + y, [re[1] for re in res]) /cls_num #计算mAP
print("mAP = ", mAp)