Dual similarity pre-training and domain difference encouragement learning for vehicle re-identification in the wild

Abstract

Existing unsupervised domain adaptation (UDA) tasks require extensive annotated wild data in the source domain to be generalized to the target domain. Additionally, the large gap between the source and target domains hinder the clustering performance. Our work concentrates on few-shot UDA task to train a robust Re-ID model from practical vehicle re-identification (Re-ID). That is to say, this task learns discriminative representations from a few labeled source data to unlabeled target data. In this paper, a novel progressive few-shot UDA learning framework for vehicle Re-ID is proposed, which consists of two branches. In source branch, the dual prior model is used to gain the color and IDs in unlabeled source data. A dual constraint label smoothing regularization (DCLSR) loss is designed to supervise extensive unlabeled source data during pre-training phase, which considers color and ID constraints to mine the similarity between unlabeled source data and a few labeled ones. The target branch develops a progressive domain difference encouragement learning method to optimize the cross-domain capability of Re-ID model. The domain difference penalty term (DDPT) is encoded by the feature variations before and after style transfer, which improves clustering results and refines the pseudo label. Comprehensive experimental results verify that the proposed approach outperforms other ones in the practical UDA task.