Why do you assume that neural network would be unable to understand (or classify) 'Gestalt' of the picture? I would not make those assumptions about perspective, because in the end ml could make 3d model of the picture and KNOW the exact perspective.