Context-aware reliability exploration for unsupervised domain adaptive person re-identification

Unsupervised domain adaptive (UDA) person re-identification (ReID) aims to transfer the knowledge from labeled source domain to unlabeled target domain, where discriminative feature representation need to be learned. Existing cluster-based pipeline generates pseudo labels as supervision for target domain training, which suffers a lot from the problem of pseudo-label noise and brings limited performance improvement. It is particularly important to note that these current methods exhibit susceptibility to the samples of different identities but with similar appearances, resulting in the label noise issue. In this paper, we propose a Context-Aware Reliability Exploration (CARE) framework that reduces the label noise via refining the target-domain pseudo labels and mining the assistance of the labeled source domain. Specifically, we propose a neighbor-based reliable supervision selection strategy to pick up samples with reliable pseudo-labels. Afterwards, we employ consistency learning in a self-supervised manner for the remaining samples with unreliable labels to capture internal discriminative information of target domain. Moreover, the knowledge transfer module assisted by the source domain dynamically aids in model learning based on the distance to the target domain, thereby minimizing the inter-domain distance for enhanced knowledge transfer. Extensive experiments demonstrate the effectiveness of our proposed method, which outperforms most existing state-of-the-art approaches on the unsupervised domain adaptive person ReID. The code is available at https://github.com/code9876ll/CARE.