Network based on epoch-to-epoch adaptation weighted triplet loss for person re-identification

Abstract

Abstract Person re-identification (Re-ID) aims to use computer vision techniques to determine whether a specific person is present in a set of images. With the widespread use of deep learning, how to efficiently make the deep neural network for person Re-ID achieve excellent performance has gained wide attention. In this paper, we propose a metric learning method based on a new Epoch-to-epoch Adaptation Weighted (EAW) triplet loss function. The EAW triplet loss function uses the variability and connectivity of metric information between epochs to guide the optimization direction of the network. It enhances the inter-class differentiation through the adaptive weight and margin, speeds up the convergence of the network, and improves accuracy without increasing cost. Meanwhile, to prevent the risk of overfitting due to the complex loss function, we regularly employ sample pairing to optimize the network. We conduct evaluation experiments on both Market-1501 and DukeMTMC-reID datasets. With the same network, our loss function can effectively improve the network performance. On the Market-1501, our method achieves 95.3% rank-1 accuracy and 89.2% mAP and on the DukeMTMC-reID, the mAP and rank-1 accuracy can reach 90.4% and 80.4% respectively. The experiments show that our method can effectively improve the accuracy and training efficiency.