JAC-Net: Joint learning with adaptive exploration and concise attention for unsupervised domain adaptive person re-identification

Abstract

Existing unsupervised domain adaptive (UDA) methods of person re-identification (re-ID) often use clustering to generate and optimize pseudo-labels. However, the pseudo-labels generated in this way contain noise, which is gradually amplified during the iterative process, leading to a lower recognition accuracy than for supervised methods. This paper proposes Joint Learning with Adaptive Exploration and Concise Attention Network (JAC-Net), which uses two identical networks to optimize person re-ID in an unlabeled target domain. Based on the pseudo-labels generated by clustering, JAC-Net optimizes the training network by combining a joint learning network (JLN) with a concise attention module (CAM). Inspired by the teacher-student network, JLN uses two identical networks to share knowledge for network learning, and also applies adaptive exploration learning strategies to automatically assign weights to the two identical networks and to balance the impact of the knowledge from the two networks. As a parameter-free attention module, a CAM is added to the feature map extracted in specific layers of ResNet50 without changing the high-order semantic features. Extensive experiments on the Market-1501, DukeMTMC-reID and MSMT17 datasets show that JAC-Net achieves well performance and reaches a similar level of supervised learning.