The receptive field is a kind of thing, from the angle of CNN visualization, is the output featuremap a node response to the input image of the area is to feel wild.
For example, if our first layer is a 3*3 core, then each node in the Featuremap that we get through this convolution is derived from this 3*3 convolution core and the 3*3 region of the original image, then we call this Featuremap node to feel the wild size 3*3
If you go through the pooling layer, assuming that the stride of the convolution layer is the 1,pooling layer size 2*2,stride is 2, then the pooling layer node is 5*5
There are a few points to note is that padding does not affect the feeling of the wild, stride only affect the next layer of featuremap feeling field, size affects the level of the feeling of the wild.
As for how to calculate the feeling field, my suggestion is the way of top to down. Let me take an example to calculate
An output of pool3 corresponds to a pool3 input size of 2*2
And so on, the corresponding conv4 input is 5*5, because 2*2 each corner plus a 3*3 convolution kernel, it is 5*5, of course, this is in the case of stride=1, but is generally stride=1, otherwise unreasonable
The input for the corresponding Conv3 is 7*7
The input for the corresponding pool2 is 14*14
The input for the corresponding Conv2 is 16*16
The input for the corresponding Pool1 is 32*32
The input for the corresponding CONV1 is 34*34
So Pool3 's feeling of the wild size is 34*34
Copyright NOTICE: This article for Bo Master original article, without Bo Master permission not reproduced.
The calculation of the sensing field in CNN