Group sparsity based sparse coding for region covariances

In the recent years, there has been an increasing interest in using sparse representations for image processing and computer vision. The main reason behind their popularity is that they could provide a more robust and efficient way of reconstructing a target by means of a limited number of atoms in a dictionary. The common practice in sparse coding is to use dictionary atoms which live in Euclidean spaces. In recent years, some studies proposed to use region covariance based dictionary atoms to come up with more effective sparse coding schemes. The optimization schemes suggested by these studies are fundamentally different than those of the standard methods since covariance matrices live in a special Riemannian manifold. In this study, we propose to enrich such a sparse coding scheme proposed by Sivalingram et al. with a group sparsity constraint. The experimental results on a face recognition task reveals that considering group sparsity improves the recognition rate.