Learning Object Representation Form Lighting Variations

Abstract

Realistic representation of objects requires models which can synthesize the image of an object under all possible viewing conditions. We propose to learn these models from examples. Methods for learning surface geometry and albedo from one or more images under fixed posed and varying lighting conditions are described. Singular value decomposition (SVD) is used to determine shape, albedo, and lighting conditions up to an unknown 3×3 matrix, which is sufficient for recognition. The use of class-specific knowledge and the integrability constraint to determine this matrix is explored. We show that when the integrability constraint is applied to objects with varying albedo it leads to an ambiguity in depth estimation similar to the bas relief ambiguity. The integrability constraint, however, is useful for resolving ambiguities which arise in current photometric theories.