Yolov3 & yolov5 output result description

This article uses yolov3 and yolov5 Engineering documents are github On ultralytics be based on pytorch Of v3 and v5 Code , The output results of the training set are basically the same , The output results are mainly introduced , This is a study note

This article uses yolov3 Code github Download address ：yolov3
See this blog for specific steps of model training ：
yolov3 model training —— Use yolov3 Train your model

This article uses yolov5 Code github Download address ：yolov5
See this blog for specific steps of model training ：
yolov5 model training ——— Use yolov5 Train your own dataset

yolov3 The output results of model training are shown in the figure below ：

yolov5 The output results of model training are shown in the figure below ：

（ This version of v3 and v5 The output result type looks exactly the same , I used the same data set for training , So the output visualization results are also very similar ,v5 The result screenshot on the server was not saved , This is used here. windows Lower view ）

1. confusion_matrix.png

confusion_matrix.png Refers to the confusion matrix （Confusion Matrix）

In machine learning , Confusion matrix （Confusion Matrix）, Also known as possibility matrix or error matrix . Confusion matrix is a visualization tool , Especially for supervised learning , In unsupervised learning, it's called matching matrix . In the evaluation of image accuracy , It is mainly used to compare the classification results with the actual measured values , The accuracy of the classification results can be displayed in a confusion matrix .
Each column of the confusion matrix represents the prediction category , The total number of each column represents the number of data predicted for that category ;
Each row represents the true category of the data , The total number of data in each row represents the number of data instances in that category ; The values in each column represent the number of real data predicted to be of this class .

The confusion matrix structure is shown in the figure below ：

2. F1_curve.png

F1_Score： The mathematical definition is F1 fraction （F1-Score）, Also known as balance F fraction （Balanced Score）, It is defined as the harmonic average of accuracy and recall . stay β=1 The situation of ,F1-Score The value range of is 0 To 1,1 It's the best ,0 Is the worst . The calculation formula is shown in the figure below ：

3. labels.jpg

The first picture classes: The amount of data in each category

The second picture labels： label

The third picture center xy

The fourth picture labels The length and width of the label

4. P_curve.png

Accuracy Precision And confidence confidence Diagram for

5. PR_curve.png

PR In the curve P It stands for Precision（ Accuracy ）,R It stands for Recall（ Recall rate ）, It represents the relationship between accuracy and recall , In general , take recall Set to abscissa ,precision Set to ordinate .PR The area enclosed under the curve is AP, All categories AP The average value is map.

If PR One of the curves in Figure A Completely enclose the curve of another learner B, It can be asserted that A Better performance than B, When A and B When crossing occurs , It can be compared according to the area under the curve . General training results mainly observe the fluctuation of accuracy and recall rate （ If the fluctuation is not great, the training effect is better ）.
Precision（ Accuracy ） Indicates the percentage of positive samples detected as correct by the classifier in all positive samples predicted , That is, in the prediction box of the current traverse , The proportion of correct target prediction frame detected . The formula is shown in the figure below ：

Recall（ Recall rate ） Indicates the percentage of all positive samples detected as correct by the classifier , That is, in all truth bounding boxes , Proportion of correct positive sample bounding box detected . The formula is shown in the figure below ：

Precision and Recall Generally, it is a pair of contradictory performance metrics ;
Improve Precision Improve the threshold of positive case prediction by two classifiers , Make the positive example predicted by the two classifiers as real as possible ;
Improve Recall Reduce the threshold of positive case prediction by two classifiers , Make the two classifiers choose the real positive examples as much as possible

6. R_curve.png

Recall rate Recall And confidence confidence The relationship between

7. results.png

（1）yolov3 And yolov5 The loss function of can be divided into three parts ： Class loss function （Classification loss）、 Confidence loss function （Confidence loss） And position loss function （Localization loss）

Box_loss：Box by CIoU The loss function means , The smaller the box, the more accurate ;
Objectness_loss：Objectness Is the average loss of target detection , The smaller the target, the more accurate the detection ;
Classification_loss：Classification Is the average loss of classification , The smaller the size, the more accurate the classification ;

val Box_loss: Verification set bounding box Loss ;
val Objectness_loss： Verification set target detection loss mean value ;
val classification_loss： Validation set classification loss mean value ;

（2）
Precision： The change curve of accuracy rate with the number of training
Recall： The change curve of recall rate with the number of training

（3）AP Value is an important evaluation index to measure the performance of target detection model classifier ,AP The higher the value, the better the classifier performance , The smaller the size, the worse the classifier performance ;AP The size of the value is equal to P-R The area of the area enclosed by the curve and the coordinate axis .
mAP Represents all label categories AP The average of the values ,mAP The bigger the value is. , It shows that the better the performance of the model .

[email protected]： It means that IoU The threshold for 0.5 At the time of the mAP The curve of value change
[email protected]:0.95： It means that IoU Threshold with 0.05 The step size is from 0.5 To 0.95 When changing mAP The curve of value change , namely IoU Threshold value 0.5、0.55、0.6、0.65、0.7、0.75、0.8、0.85、0.9、0.95 Average at mAP.

8. results.txt

results.txt The last three columns in are the validation set results , The previous is the result of the training set , All columns are ：
Training times ,GPU Consume , Bounding box loss , Target detection loss , Classified loss ,total,targets, Picture size ,P,R,[email protected], [email protected]:.95, Verification set val Box, Verification set val obj, Verification set val cls

test_batch0_labels Pictures like this are the visual results of model training output , I won't repeat ~