Youtube https://www.youtube.com/watch?v=44tlnmmt3h0 Key background and foreground training process with obj/no obj really suffer by hilghy imblanced classfication problem to solve that, use focal loss They also called the network RetinaNet Paper Focal Loss for Dense Object Detection, 3426+ implementation tf - been tested, 15+ star tf, Keras only 3lines of code 255+ stars Some Hairy details 为什么 YOLOv3 用了 Focal Loss 后 mAP 反而掉了？