Graph Regularized Double Log-determinant Minimization for Subspace Clustering

Subspace clustering methods have been successful in various applications such as face clustering. Existing methods usually adopt the l_1 or l_2-based norms to minimize the fitting error. However, the fitting residual constructed with such norms usually has structural information since these norms are computed in an element-wisely independent way and all elements of the residual are treat-ed equally. To maximumly minimize structural information in residual so as to maximumly retain such information in recovered data, in this paper we propose to minimize fitting error of each example with log-determinant rank approxima-tion. Extensive experiments verify the effectiveness of the proposed method.