Quasi-invariant illumination recognition for appearance-based models, taking advantage of manifold information and non-uniform sampling

Abstract

The appearance of an object can be greatly affected by the changes in illumination conditions in the scene where that object is seen. In this paper, we propose a novel method for building appearance-based models that are tolerant to lighting variations. This new approach is based on a Non-Uniform Sampling technique, so that we use a very small number of images for generating those models, and this reduces the computing time and storage space required for the modeling stage, with respect to Uniform Sampling techniques. We have tested the proposed algorithm with a comprehensive set of objects with different appearance, and the high recognition rate obtained in these experiments shows the effectiveness of the proposed method. Also, the necessary time to compute the associated manifold in eigenspace is dramatically reduced. © 2009 Springer-Verlag Berlin Heidelberg.