Heterogeneous dual network with feature consistency for domain adaptation person re-identification

To reduce the noisy pseudo-labels generated by clustering for unsupervised domain adaptation (UDA) person re-identification (re-ID), the method of collaborative training between dual networks has been proposed and proved to be effective. However, most of these methods ignore the coupling problem between dual networks with the same architecture, which makes them inevitably share a high similarity and lack heterogeneity and complementarity. In this paper, we propose a heterogeneous dual network (HDNet) framework with two asymmetric networks, one of which applies convolution with limited receptive fields to obtain local information and the other uses Transformer to capture long-range dependency. Additionally, we propose feature consistency loss (FCL) that does not rely on pseudo-labels. FCL focuses more on the consistency of the sample in the feature space rather than the class prediction space, driving the feature learning of UDA re-ID from the task level to the feature level. Furthermore, we propose an adaptive channel mutual-aware (ACMA) module which contains two branches to focus on the global and local information between channels. We evaluate our proposed method on three popular datasets: DukeMTMC-reID, Market-1501 and MSMT17. Extensive experimental results demonstrate that our method achieves a competitive performance.