If it is possible to interpret an image as a projection of rectangular forms, there is a strong tendency for people to do so. In effect, a mathematical basis for a vector space appropriate to the world, rather than to the image, is selected. A computational solution to this problem is presented. It works by backprojecting image features into three-dimensional space, thereby generating (potentially) all possible interpretations, and by selecting those which are maximally orthogonal. In general, two solutions that correspond to perceptual reversals are found. The problem of choosing one of these is related to the knowledge of verticality. A measure of consistency of image features with a hypothetical solution is defined. In conclusion, the model supports an information-theoretic interpretation of the Gestalt view of perception.