[vivid understanding] the meanings of various evaluation indicators commonly used in deep learning TP, FP, TN, FN, IOU and accuracy

Whether for the classification of the image field 、 Segmentation problem , Or object classification in point cloud field 、 Scene segmentation , When evaluating a model , Often involves TP、FP、TN、FN、IoU、Accuracy... And other indicators , So what do these indicators mean ？ I have Mr. duck 、 The monkey guest and the panda steward come to show you .

start

Why ！ Find out 10 A lovely little animal , But we don't know what animal it is , You don't have to study deeply to classify ？

At this time, we just take out Train well Of Model , Throw these little animals in , After a long wait , We have the following prediction results ：

Obviously , The model does not complete the task perfectly . But how to evaluate it ？

Basic indicators ：TP、FP、TN、FN

To some Category A Speaking of ：T = true, Indicates correctly classified ;F = false, For error classification ;P = Positive, Indicates that the classification result is A;N = Negative, Indicates that the classification result is non A.

• TP（True Positive）： Correct and divided into A Number of （ the truth is that A）.
• FP（False Positive）： Misclassification A Number of （ the truth is that Not A）.
• TN（True Negative）： Correct and divided into Not A Number of （ the truth is that Not A）.
• FN（False Negative）： Misclassification Not A Number of （ the truth is that A）.

that ： For the above 3 In terms of animal categories ：

Category	TP	FP	TN	FN
The duck	5	1	3	1
monkey	1	1	7	1
Panda	2	0	8	0

You can see , For each category ,TP+FP+TN+FN == 10（ common 10 Animals ）

Accuracy rate ：Accuracy

meaning ： Accuracy of classification of all small animals , in total 10 A little animal , among 8 The categories are correct .
Calculation ：8 / 10 = 80.00%

Accuracy ：Precision

meaning ： In the prediction results of each category , Proportion belonging to the correct classification
Calculation ：P = TP / (TP + FP)

• The duck ：5 / （5 + 1）= 83.33%
• monkey ：1 / （1 + 1）= 50.00%
• Panda ：2 / （2 + 0）= 100.00%

Occurring simultaneously than ：IoU

meaning ： The ratio of the intersection and union of the predicted result and the real value of a certain class

• The duck ： True value （6 A duck ） And Predicted results （5 A duck +1 Only monkey ）, Its intersection is 5 only , Union as 7 only .
  Calculation ：IoU = 5 / 7= 0.714
• monkey ： True value （2 Only monkey ） And Predicted results （1 A duck +1 Only monkey ）, Its intersection is 1 only , Union as 3 only .
  Calculation （ monkey ）：IoU = 1 / 3= 0.333
• Panda ： True value （2 Bear ） And Predicted results （2 Cats ）, Its intersection is 2 only , Union as 2 only .
  Calculation （ Panda ）：IoU = 2 / 2= 1

so ：IoU The value is approximately close to 1, Indicates that the better the prediction result of this class

The average ratio of crossing and merging ：mIoU

meaning ： All categories of IoU Sum up , Divide by the number of categories
Calculation ：（0.714 + 0.333 + 1）/ 3 = 0.682