Cross-domain Person Re-identification Based on the Sample Relation Guidance

Yue Zhang, Fanghui Zhang, Shichao Kan, Linna Zhang, Jiaping Zong, Yigang Cen

Abstract: The existing mainstream cross-domain person re-identification (Re-ID) methods mainly focus on reducing the deviation of the generated pseudo labels, and they did not introduce veracious label information for algorithm training on the unlabeled target domain. In this paper, we propose a new sample relation guidance (SRG) method. Specifically, the sample relation is a real label, which represents a definite positive sample pairs’ relation or negative sample pairs’ relation. Here, we construct a triple-branch network to form sample relation labels to improve the expressive power of features. In addition, the potential relationship of target domain label loss and source domain label loss is explored, and an adaptive adjustment label loss (ADLL) method is proposed, which effectively improves the generalization performance of the model. Extensive experiments over three benchmarks proved that our method outperforms the state-of-the-art methods.