Rank constrained recognition under unknown illuminations

Recognition under illumination variations is a challenging problem. The key is to successfully separate the illumination source from the observed appearance. Once separated, what remains is invariant to illuminant and appropriate for recognition. Most current efforts employ a Lambertian reflectance model with varying albedo field ignoring both attached and cast shadows, but restrict themselves by using object-specific samples, which undesirably deprives them of recognizing new objects not in the training samples. Using rank constraints on the albedo and the surface normal, we accomplish illumination separation in a more general setting, e.g., with class-specific samples via a factorization approach. In addition, we handle shadows (both attached and cast ones) by treating them as missing values, and resolve the ambiguities in the factorization method by enforcing integrability. As far as recognition is concerned, a bootstrap set which is just a collection of 2D image observations can be utilized to avoid the explicit requirement that 3D information be available. Our approaches produce good recognition results as shown in our experiments using the PIE database.