Network Improved by Auxiliary Part Features for Person Re-identification

Abstract

Person re-identification (ReID) is an important issue in computer vision area. It focuses on identifying people under different scenarios. In this paper, we test the contributions of local part features in ReID system. With the auxiliary of local part features, our model achieves significantly improvements, which achieves rank-1 accuracy of 91.7% on market1501 dataset and 82.6% on MARS dataset. We also test the feasibility of using densenet as backbone model in ReID system. With densenet as our backbone model, our method achieves state-of-art performance and simultaneously reduces the model size enormously.