As far as I understand, segmentation treats each pixel separately from all the others. Thus, if one would randomly shuffle all the pixels in an image (and map the mask accordingly of course), it would not matter to the network.
Effectively, that means the only feature of a pixel being predicted upon its color.
Do you agree?
However, I am interested in segmentation based on the texture of a surface as well, not only upon its colors.
I thought one way this could be done is to segment (1) an image (2) a low-resolution version of the image. (3) combine manually the predictions of the two.
The low-resolution version of the image de-facto averages many adjacent pixels. Thus making predictions for the low-resolution picture would take into account automatically values of adjacent pixels as well. Taking into account values of adjacent pixels is tantamount to looking at the texture.
Any thoughts?