Adversarial Style-Irrelevant Feature Learning With Refined Soft Pseudo Labels for Domain-Adaptive Vehicle Re-Identification

Domain-adaptive vehicle re-identification is a challenging task that aims to transfer the knowledge from a labeled source domain to an unlabeled target domain for effective vehicle re-identification. Two key challenges hinder the effectiveness of existing domain-adaptive methods that rely on clustering algorithms: 1) generation of false pseudo labels due to clustering errors and 2) significant distributional disparities between the source and target domains that can mislead the learning process. This study tackles these two challenges with an adversarial style-irrelevant feature learning framework guided by refined soft pseudo labels. This framework not only enhances the confidence level of pseudo labels but also mitigates domain distributional disparities, leading to improved domain-adaptive capabilities. Specifically, a centroid similarity ranking is utilized to refine soft pseudo labels. This process preserves knowledge from hard-to-distinguish categories while discarding the easy-to-distinguish ones. This approach strengthens learning from more difficult samples. In addition, with the guidance of the refined soft pseudo labels, an adversarial domain style feature branch is introduced. This branch aims to reduce the style information contained in vehicle features and narrow the distributional disparities across domains, thereby improving the domain-adaptive capability of the vehicle re-identification model. Extensive experiments demonstrate the superiority of our proposed method over state-of-the-art methods in cross-domain vehicle re-identification.