kkkkkkkkq的博客

Rich feature hierarchies for accurate object detection and semantic segmentation论文解读http://blog.csdn.net/hjimce/article/details/50187029讲得很好，下面再加一点自己的理解。1、在测试阶段，每个建议框通过CNN获取特征，然后分别用训练好的判断每个类别的SV

http://blog.csdn.net/hunterlew/article/details/71075925 部分说的很好。RPN训练设置：根据ANCHOR_SCALES和ANCHOR_RATIOS得到(1)width/RPN_FEAT_STRIDE*height/RPN_FEAT_STRIDE*len(ANCHOR_SCALES)*len(ANCHOR_RATIOS)个anch