1、loss不下降:
https://blog.csdn.net/qq_19329785/article/details/84260201
https://blog.csdn.net/zongza/article/details/89185852
https://blog.csdn.net/liuweiyuxiang/article/details/80856991
2、检测中faster-rcnn的loss是nan:大部分是坐标越界造成的;
https://blog.csdn.net/qq_14839543/article/details/72900863
3、检测中faster-rcnn的bbox_loss为0,主要是因为fg太少,遇到过第一阶段的rois中,假设batch_size是128,得到fg/bg=1/127,很可能bbox_loss等于0,因为rois中fg就是1,在第二阶段计算bbox_loss时只计算正样本,而正样本就1个,因此loss很小,为0;
https://blog.csdn.net/u011070171/article/details/70256611
https://blog.csdn.net/KevinLee200204/article/details/79041586