网络训练

1、hard mining

         这个trick来自检测算法DPM论文，解决2个问题：

         1）正、负样本失衡问题。

论文第4节介绍了hardmining的具体方法，顺带介绍了这个问题。“When training a model for object detection we often have a very large number of negative examples (a single image can yield 105 examples for a scanning window classifier). This can make it infeasible to consider all negative examples simultaneously. Instead, it is common to construct training data consisting of the positive in- stances and “hard negative” instances.”

       2）过多的简单样本问题，导致训练效果不理想。

这个问题可以从下图svm的超平面去理解，红色的正、负样本最有效，可以让模型学习得到最佳的分解线，而黑色的正负样本就是比较冗余的简单样本，从中抽样一部分数据即可。