Projective Weight-Based Unsupervised Laplacian Graph Learning for Person Re-Identification

For unsupervised person re-identification, traditional Laplacian regularisation based dictionary learning methods encounter the fixed weight problem and thus impair the match rate. To address this limitation, we develop a novel unsupervised dictionary learning approach to learn a discriminative representation. The proposed approach takes the efficiency of $l_{2}$ graph regularization with a closed-form solution into account. Our approaches achieve very promising results on the challenging VIPeR dataset.