In this paper, we propose the Compound Mutual Subspace Method (CPMSM) as a theoretical extension of the Mutual Subspace Method, which can efficiently handle multiple sets of patterns by representing them as subspaces. The proposed method is based on the observation that there are two types of subspace perturbations. One type is due to variations within a class and is therefore defined as "within-class subspace". The other type, named "between-class subspace", is characterized by differences between two classes. Our key idea for CPMSM is to suppress within-class subspace perturbations while emphasizing between-class subspace perturbations in measuring the similarity between two subspaces. The validity of CPMSM is demonstrated through an evaluation experiment using face images from the public database VidTIMIT.
展开▼