Dual Attention Network for Unsupervised Domain Adaptive Person Re-Identification

Clustering-based unsupervised domain adaptive person re-identification methods reduce much of the annotation cost. However, many pseudo-labels with inaccurate labels are produced when fine-tuning the target domain as a result of the shortcomings of the less than ideal network model and less than excellent clustering. In order to remove the mislabeled pseudo-labels and improve the purity of the network to obtain the correctly labeled pseudo-labels, this paper proposes a dual attention network framework. Specifically, in order to make the network learn to focus on the focal object, this paper introduces the CBAM attention mechanism, which can focus well on the information on the image channel and space, and then identify the correct and incorrect pseudo-labels to improve the overall performance of the network model. In addition, in order to allow the network to extract richer semantic information, this paper introduces Non-local blocks, which can directly capture the remote dependencies between image features by computing the interaction between any two locations. Accordingly, we conducte extensive experiments on four common unsupervised domain adaptive person re-identification tasks, namely, DukeMTMC?Market-1501, Market-1501?DukeMTMC, DukeMTMC?MSMT17, and Market-1501?MSMT17, in which the mAP/R1 and baseline compared to 3.6%/1.4%, 1.8%/2.5%, 4.7%/5.2%, and 3.4%/3.0%, respectively.