Inverse of extract_image_chip()

[pfeatherstone]

Is there a way to get an inverse transformation for extract_image_chip() ? Looking at the code, it's a bit more complication than I originally thought. However, shouldn't it just be a point_affine_tranforsmation (rotation + translation) + resize ? In which case, those should all be invertible with resize + inv(point_affine_tranforsmation(...)) ?

I'm trying to remap the landmark detections from a custom model computed in the extract chip, back into the original image.

[arrufat]

Does #2781 help?

[pfeatherstone]

So use get_mapping_to_chip() then possible inv() ? I'll check it out.