Tree-Guided Group Sparse Based Representation For Person Re-Identification

Most existing person re-identification (Re-ID) methods focus on two fundamental problems: feature representation and metric learning. A person representation with thousands of dimensions tends to lead to the over-fitting problem, since only hundreds of training samples are available. To solve the small sample size problem, dimensionality reduction techniques are resorted for metric learning. Before dimension reduction, however, the image acquisition and feature extraction process may introduce some noise and redundant information, which lead to unexpected loss of discriminative power. To avoid this, we consider the structural information of human body along with supervision, and then introduce the Tree-guided Group Sparse (TSG) method for feature selection before dimension reduction. Experimental results on VIPeR dataset show the improvement of performance by applying TGS to the existing metric learning methods and demonstrate TGS can effectively remove noise and redundancy in person representation.