Chart referencehttp ://blog.sina.com.cn/s/blog_9db078090102whzw.html
For example, there are 10 categories and 20 samples, and the confidence of one category of car is shown in the following table, in descending order.
It can be seen from the table gt_label that there are 6 positive examples and the others are negative examples. PASCAL VOC CHALLENGE The method for calculating AP after 2010 is: Assuming that there are M positive examples in N samples , as shown in the above table, N is 20 , M is 6 , then there are 6 recall values, which are 1/6 , 2/6 ,3/6,4/6,5/6,6/6 . For each recall value, there are many ways to take top , so there is a maximum precision among the many methods corresponding to each recall value (including the method equal to this recall), and the sum of the maximum precision corresponding to each recall is averaged i.e. AP .
For example , the recall of 2/6 , look up the table above, you can get the type of recall2/6 value: from the second to the fifth , and to the sixth in the above table , because the corresponding is a positive example, so it is not a recall of 2 /6 range (because there are already 2 positive examples in front, if you add another positive example, the recall value is 3/6), the maximum precision corresponding to these methods is 2/2 . Similarly, recall 4/6 is the range from the beginning of the fourth positive example ( 4/7 ) to the front of the fifth positive example ( 4/10 ), and the corresponding maximum price is 4/7. The following table
So mAP is the average of 10 types of APs.