Person Re-identification by Smooth Metric Learning

Person re-identification, which is to match people across areas covered by multiple non-overlapping surveillance cameras, has drawn great research interests in the video surveillance domain. Previous research mainly focus on feature extraction and distance measure, where the former aims to find robust feature representation while the latter seeks to learn an optimal metric space. Because of the introduction of supervised information, metric learning methods can usually achieve better performance over feature based methods. As one of the most representative metric learning method, the Large Margin Nearest Neighbor (LMNN) algorithm was recently applied in person re-identification task and achieved satisfactory results [1]. However, LMNN uses a standard hinge loss function, which is neither differentiable everywhere nor time-efficiency due to the usage of all training samples. In this paper, we propose to replace hinge loss function with the logistic loss function, which transforms LMNN to a smooth unconstrained convex optimization problem easily solved with gradient descent algorithm. Thereafter, we further design a stochastic sampling scheme to accelerate the optimization process of the above problem with randomly selected training samples. Extensive comparative experiments conducted on two standard datasets have shown the effectiveness and efficiency of the proposed algorithm over a series of standard baseline methods.