Exploiting matching local information for person re-identification

Person re-identification task with the main aim is to associate the instances of the same person captured by different cameras in a surveillance camera network usually employs the detection results. As a consequence, misalignment of detected bounding boxes and background information are the two main factors that lead to reducing the performance of person re-identification.To tackle with these challenges, the state-of-art in person re-identification methods proposed to employ attention mechanism or body parts detection. However, these methods have high complexity and computational cost, which can be reduced by using Earth Movers Distance (EMD) instead. Therefore, this paper formulates local matching as a distance calculation of two probability distributions and applies Earth Movers Distance (EMD) to compute the optimal matching between two sets of stripes in order to address an issue in the AlignedReID++ method. Different experiments have been conducted on both single-shot and multi-shot person re-identification. The obtained results have shown the improved performance of the proposed method compared with the baseline method. The matching rates at rank1 obtained by the proposed method are 49.59%, 83.36%, and 78.47% on VIPeR, Marketl501-Partial, and DukeMTMCReID-Partial, respectively.