空洞卷积 hole Atrous 原理+图解析+应用

2015 CVPR

Multi-Scale Context Aggregation by Dilated Convolutions

Hole

（Both large feature map & receptive field）

• hole algorithm：up-sample the original filter by a factor of the strides(rate=2)
  efficient dense sliding window feature extraction

  • standard convolution : responses at only 1/4 of the image positions
  • convolve image with a filter “with holes”:responses at all image positions

• RF and stride

按照公式，stride变小，要想保持receptive field不变，那么，就应该增大kernel size。于是就有了接下来的hole算法。

Example：

pooling layer stride