Common Structured Low-Rank Matrix Recovery for Cross-View Classification

Low-rank multi-view subspace learning (LMvSL) has been an essential solution to the problem of cross-view classification. Despite the promising performance on real applications, it still remains challenging to classify objects when there is a large discrepancy between gallery data and probe data. In this paper, we propose a Common Structured Low-rank Matrix Recovery (CSLMR) algorithm to elegantly handle view discrepancy and discriminancy simultaneously. Specifically, our CSLMR incorporates common representation constraint and structured regularization into the fundamental model of LMvSL to learn a discriminant latent subspace. Furthermore, an efficient optimization method is developed and the complexity analysis of CSLMR is presented for completeness. Experimental results on CMU PIE dataset demonstrate the superiority of our CSLMR.