Dual Pseudo Label Refinement for Unsupervised Domain Adaptive Person Re-Identification

Abstract

Although clustering-based unsupervised domain adaptive person re-identification has achieved promising progress, negative influence of noisy pseudo labels is still unsolved. Training model with noisy labeled data will mislead feature representation and ultimately affect the improvement of model performance. To tackle the above problem, we propose a dual pseudo label refinement framework for unsupervised domain adaptive person re-identification. It has two pseudo label refinement modules, one learns cross consistency of corresponding features between two collaborative networks and the other explores mutual consistency between global and local feature spaces. By working complementarily and jointly, the two modules optimize the quality of training datatset. Specifically, a mean teacher-student framework with multiple branches is designed to extract global and local features. Cross consistency of corresponding features between mean teacher and student networks is learned to select candidate samples for training. Meanwhile, mutual consistency of global and local features in mean teacher network is also explored to select other candidates. Finally, the intersection of two selected candidate datasets is used to re-train the framework. Being trained with selected reliable samples iteratively, the framework becomes more and more effective and robust. Extensive results on benchmark datasets confirm that our dual pseudo label refinement brings performance improvement and outperforms the state-of-the-art clustering-based unsupervised domain adaptive re-identification approaches.