Figure 5.

Example of extending receptive field of spatial channel-wise convolution. Taking the blue pixel as example, when only using the spatial channel-wise convolution, as shown in (A), the effective receptive field of the red pixel in the output feature map is 1, which is the corresponding blue pixel in the input feature map. When adding the traditional convolution layer before spatial channel-wise convolution (convolution kernel size = 3, padding = 1, stride = 1), as shown in (B), the effective receptive field of the blue pixel in the output feature map is 3, which is the same size of the convolution kernel in the input feature map.