Generalization abilities of appearance-based subspace face recognition algorithms

In this paper we present an efficient method to test the generalization abilities of subspace face recognition algorithms. The main motivation for this work is the lack of detailed analysis of this problem in current literature. Generalization ability of face recognition algorithm is the ability to recognize new individuals, which were not part of the training process. To illustrate our idea we used well-known recognition algorithms (PCA, ICA and LDA) and the FERET date set. Our results show that even these well-known algorithms have poor generalization abilities in some implementations.