评价指标

时隔一年,Started learning again.Hahaha not much nonsense,Let's start with some of the content I've put together recently.The knowledge points to be sorted out today are several evaluation indicators commonly used in the field of target detection,准确率Precision,召回率Recall,交并比IoU,平均精度AP,多个类别APthe average ofmAP等等.

混淆矩阵

目标检测模型通常会输出很多个检测框,我们是通过统计并计算每个检测框是否能检测到目标的各种占比来衡量模型的检测效果,因此,我们会把检测框分成如下四种情况,判断的依据主要是通过计算交并比（IoU,这个下面会讲到）.

• TP (True Positive) ,真的正样本 = 正样本 被分类为 正样本;
• TN (True Negative) ,真的负样本 = 负样本 被分类为 负样本;
• FP (False Positive) ,假的正样本 = 负样本 被分类为 正样本;
• FN (False Negative) ,假的负样本 = 正样本 被分类为 负样本（通常为漏检）;

According to the detection situation of the detection frame,我们写成一个4x4的矩阵形式（称为混淆矩阵）：

for an image,The predicted situation and the real sample situation are as follows（记住这个,Later calculations will not be confusing）

Precision 精确率（精度）

Precision：指的是在模型预测的结果中,其中正确的有多少个.看上面的图,预测结果有TP+FP.

$$Precision=\frac{TP}{TP+FP}$$

Recall 召回率

Recall：指的是在所有的真实目标中,其中正确的有多少个.看上面的图,真实目标有TP+FN.

$$Recall=\frac{TP}{TP+FN}$$

IoU 交并比

IoU：It refers to the ratio of the intersection and union of the detection frame and the rectangular frame marked by the sample.

平均精度 AP

AP（Average Precision）指的是Precision-Recall曲线下的面积.

Start with a single category（猫）开始计算,假设一共有3张图片,绿色框是GT（7个）,红色框是预测框（7个）with confidence.现在假设IOU=50%,The following table is obtained in order of confidence

After getting the statistics,根据confidence取不同的阈值,可以计算Precision和Recall

This way we can calculate the type of cat based on the tableAP值了

$$猫AP=\sum _{i=1}^{Rank}(Recal{l}_i-Recal{l}_{i-1})\times Precisio{n}_i$$

即：猫AP=(0.14-0)x1.0 + (0.28-0.14)x1.0 + (0.42-0.28)x1.0 + (0.57-0.42)x1.0 + (0.71-0.57)x0.71=0.6694

mAP

mAP (means Average Precision) 即各类别的AP值的均值

$$mAP=\sum _j^{class\_number}Class\_A{P}_j$$

总结

完结.谢谢,觉得有帮助的可以点个赞~