Collaborative Feature Learning and Credible Soft Labeling for Unsupervised Domain Adaptive Person Re-Identification

Cross-domain person ReID remains a challenging task for its difficulty in transferring knowledge from labeled source domain to unlabeled target domain. Aiming at the problem of weak interaction of cross-domain feature learning and inaccurate pseudo-label estimation in target domain, we propose a novel framework termed Collaborative Feature Learning and Credible Soft Labeling (CFSL) to achieve efficient domain adaptation for ReID. By designing a Collaborative Feature Extraction (CFE) module, a more powerful and discriminative image description is generated. Specifically, CFE jointly learn robust features by integrating both global and local clues on two domains and mining both cross-domain invariant features and domain-specific features. Moreover, we exploit a Dual Soft Labeling (DSL) strategy in target branch to obtain more credible and reliable identity estimations. Experimental results demonstrate the effectiveness of our method and show significant performance improvements over state-of-the-art methods on two public benchmarks.