How to implement Image Pooling in pytorch

Hi I am working on Implementing DeepLabV3, I would like to know how to implement Image Pooling in pytorch.

What exactly do you mean by “image pooling”?

Atrous convolution is the same as dilated convolutions.

Look here, you need to set different values for the dilation.

Hi richard,
I do not exactly know what is Image pooling But since they are concatenating all the layers obtained by spatial pyramid pooling later so I am expecting It is a pooling operation which is not changing the dimensions. I was guessing It will decrease the number of channels(I might be wrong). I have gone through other architecture DeepLabV3+ Encoder-Decoder with Atrous Separable Convolution for semantic Image segmentation. Even there no mention about Image Pooling

Hi ignacio-rocco,
I did not understand the Image pooling method and how to Implement it?

It seems that you want to implement ASPP.
From Figs 2d and 5 it seems that this means running a set of convolutional layers to obtain a sense feature map, and then applying a set of different convolutions with different dilations. The results are then concatenated. This feature is a multi-scale representation of the image. (Somehow also similar to hypercolumns)

I think you are trying to implement the image-level features. The image-level features are exploited in ParseNet and it is implemented by a global average pooling. You can confirm it (section 3.3), but I almost sure about that. I will start to do the same work applied to COCO2017 dataset for semantic segmentation.

1 Like