An Improved Person Re-Identification Method based on AlignedReID ++ algorithm

Abstract

Person re-identification (ReID) is a popular research topic in computer vision. It focuses on matching a given person from an image dataset captured by many non-overlapping cameras. It remains challenging duo to the influences of pose, illumination, occlusion, and background confusion. In this paper, an improved ReID approach based on the AlignedReID ++ algorithm is proposed. Three effective training tricks are introduced to improve the effectiveness of the AlignedReID ++ algorithm. Training loss, accuracy, and mean average precision (mAP) are used as measure metrics. Extensive experiments are implemented on the ResNet50 and DenseNet121 backbone networks. Our implementation gains the Rank-1 accuracy and mAP of 93.7% and 91.2%, respectively. The source code of the improved AlignReID ++ method is available on request.