Joint transfer component analysis and metric learning for person re-identification

Abstract

A novel and efficient metric learning strategy for person re-identification is proposed. Person re-identification is formulated as a multi-domain learning problem. The assumption that the feature distributions from different camera views are the same is overthrown in this Letter. ID-based transfer component analysis (IDB-TCA) is proposed to learn a shared subspace, in which the differences in the feature distribution between source domain and target domain are significantly reduced. Experimental evaluation on the CUHK01 dataset demonstrates that metric learning with IDB-TCA embedded outperforms state-of-art metric methods for person re-identification.