When looking at the two functions GeomNorm.process and extrapolate_mask for color images, two different mask types are required:

• GeomNorm takes a 3D mask: https://gitlab.idiap.ch/bob/bob.ip.base/-/blob/a9109e06264f2e28ece3b131da9babb0365b6f36/bob/ip/base/include/bob.ip.base/GeomNorm.h#L99 (implementation here: https://gitlab.idiap.ch/bob/bob.ip.base/-/blob/a9109e06264f2e28ece3b131da9babb0365b6f36/bob/ip/base/include/bob.ip.base/GeomNorm.h#L166)
• extrapolate_mask takes a 2D mask: https://gitlab.idiap.ch/bob/bob.ip.base/-/blob/a9109e06264f2e28ece3b131da9babb0365b6f36/bob/ip/base/include/bob.ip.base/Affine.h#L445

This is a bit inconsequent. I am not sure, why GeomNorm requires a 3D mask for color images. Each pixel will have the same visibility independent of the color channel. Thus, I would propose to reduce the mask for GeomNorm to 2D. We can also have the possibility to have both 2D and 3D masks, in case you require backward compatibility (which I could understand).