Kernelized Relaxed Margin Components Analysis for Person Re-identification

Person re-identification across disjoint camera views plays a significant role in video surveillance. Several margin-based metric learning algorithms have recently been proposed to learn an optimal metric, with the goal that samples of the same person always belong to the same class while those from different classes are separated by a large margin. These approaches require no modification or extension in order to solve problems of multiple (as opposed to binary) classification. However, the formation of the margin in these methods is not scalable, and thus cannot adequately use inter-class information according to the relevant practical application. To address this issue, we propose a novel algorithm called Relaxed Margin Components Analysis (RMCA) to "relax" the margin constraint. Furthermore, we equip our RMCA with a kernel function to form a Kernelized RMCA (KRMCA) to learn non-linear distance metrics in order to further improve re-identification accuracy. Promising results from experiments on several public datasets demonstrate the effectiveness of our method.