The meaning of yolov5 training visualization index

background

stay v5 During the training , Use wandb Is a very common visualization tool , It has many charts , It can visualize the evaluation indexes that can be extracted in the training process , Do a clean-up today .

Visual indicators

LOSS

loss It is divided into cls_loss, box_loss, obj_loss In the third part of .
cls_loss be used for Supervision category classification , Calculate whether the anchor frame and the corresponding calibration classification are correct .
box_loss be used for Supervise the regression of the detection box , Error between prediction frame and calibration frame （CIoU）.
obj_loss Used to supervise grid Whether there is an object in , Calculate the confidence of the network .
Reference material ：https://blog.csdn.net/u011994454/article/details/119564834

metrics

mAP([email protected]), This is a higher standard for detection capability .
mAP([email protected]), Follow Pascal VOC mAP The standard calculation method is consistent ;

mAP([email protected][0.5:0.05:0.95]), Need to compute 10 individual IoU Under the threshold mAP, Then calculate the average . This evaluation index is better than only considering general IoU threshold (0.5) The evaluation index can better reflect the accuracy of the model .

https://www.jianshu.com/p/fd9b1e89f983