Yolov5 analysis | parameters and performance indicators

The ginseng

• conf_thres：Confidence Threshold, Confidence threshold , That is, the value on the following picture . Only the prediction probability exceeds conf_thres Forecast results of .

• iou_thres：Intersect over Union Threshold, Intersection ratio threshold .
  • IOU value ： Prediction box size ∩ True frame size / Prediction box size ∪ True frame size . The value of intersection and union of prediction frame and real frame .
  • iou_thres stay detect.py in ：
    • The bigger it is , It is easy to predict different results for the same item As a Multiple predictions for multiple items , Lead to multiple prediction results for an item .
    • The smaller it is , It is easy to predict different results for multiple items As a Different predictions for the same item , As a result, there is only one prediction result for multiple items .

Evaluation parameters

• P（Precision）, Accuracy , Antitype A Come on （ All of the following are predicted to be A Of ）：
  • P = Correct number / The total number of predictions
  • or P = Correct number / Correct number + Wrong number
  • namely , What percentage of things predicted are correct .
• R（Recall）, Recall rate , Antitype A Come on （ All of the following are classified as A Class ）：
  • R = Predict the right number / real A Total number of classes .
  • or R = Predict the right number / Predicted A + Unforeseen A
  • namely , What percentage of predicted things are found .
• F1（F-Measure）,F value ,P And R The harmonic average of
  • F=1/(λ1P+(1−λ)1R),λ = 0.5 when , Simplified as F1
  • F1 = 2PR / (P + R)

example ： For a dichotomous problem ：
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
To classify animals , It is divided into cat and non cat . Sample , The cat has 350 Share , Non cat 150 Share . Predicted 400 Share is considered a cat , among , The right ones 300 Share , The mistakes are 100 Share . be ：P = 300 / 400 = 0.75、R = 300 / 350 ≈ 0.86

• PR chart ： Abscissa is R value , The ordinate is P value , The curve shows that when the recall rate is R when , Accuracy P Size .

example ： Suppose the number of pictures in the test set is 100 , And tested a hundred .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

• When A Class R Value to 0 when , Means with a certain accuracy P1 Calculation , Not detected A class .
• When A Class R Value to 0.1 when , Means with a certain accuracy P2 Calculation , Detected. A Of the total number of classes 10%.
  . . . . . . .

And so on , When A Class R Value to 1 when , Means with a certain accuracy P11 Calculation , All... Are detected A Class label . It can be seen that , along with R Increase in value ,A The predicted number will also increase , That is, those with low probability A Classes are gradually predicted . therefore , Can be judged P The value is as follows R To rise and fall .P Value if set to 100, That is, the marked area must belong to A class , This probability is very low , So maybe a A Classes are not recognized , Leading to R The value is 0. And if P Value to 0.1, The marked area is required to have a 10% probability of belonging to A The class can , Then there will be many A Class is marked ,R The value is also further improved . therefore ,PR The larger the area at the bottom left of the figure , It means that the effect of the model on the data set is better .

• AP（Average Precision）： Calculate in some way AP Value to express PR The area of a graph

• mAP（mean Average Precision）： Of each class AP The average of the values . Used to express the performance of multi class label prediction , Such as AP equally ,mAP The higher the , The better the performance .

  • [email protected]： When IoU by 0.5 At the time of the mAP.
  • [email protected] : .95： When IoU by range(0.5 : 0.95 : 0.05) At the time of the mAP The average of .