Progressive learning with multi-scale attention network for cross-domain vehicle re-identification

Abstract

Vehicle re-identification (reID) aims to identify vehicles across different cameras that have non-overlapping views. Most existing vehicle reID approaches train the reID model with well-labeled datasets via a supervised manner, which inevitably causes a severe drop in performance when tested in an unknown domain. Moreover, these supervised approaches require full annotations, which is limiting owing to the amount of unlabeled data. Therefore, with the aim of addressing the aforementioned problems, unsupervised vehicle reID models have attracted considerable attention. It always adopts domain adaptation to transfer discriminative information from supervised domains to unsupervised ones. In this paper, a novel progressive learning method with a multi-scale fusion network is proposed, named PLM, for vehicle reID in the unknown domain, which directly exploits inference from the available abundant data without any annotations. For PLM, a domain adaptation module is employed to smooth the domain bias, which generates images with similar data distribution to unlabeled target domain as “pseudo target samples”. Furthermore, to better exploit the distinct features of vehicle images in the unknown domain, a multi-scale attention network is proposed to train the reID model with the “pseudo target samples” and unlabeled samples; this network embeds low-layer texture features with high-level semantic features to train the reID model. Moreover, a weighted label smoothing (WLS) loss is proposed, which considers the distance between samples and different clusters to balance the confidence of pseudo labels in the feature learning module. Extensive experiments are carried out to verify that our proposed PLM achieves excellent performance on several benchmark datasets.