Semi-supervised pedestrian re-identification via a teacher–student model with similarity-preserving generative adversarial networks

Abstract This paper describes a pedestrian re-identification algorithm, which was developed by integrating semi-supervised learning and similarity-preserving generative adversarial networks (SPGAN). The task aimed to rapidly capture the same target using different cameras. Importantly, this process can be applied in field of security. Because real-life environments are complex, number detected identities is uncertain, cost manual labeling high; therefore, it difficult apply model based on supervised scenarios. To use existing labeled dataset large amount unlabeled data application environment, report proposes model, combines teacher–student with SPGAN. SPGAN used reduce difference between domain source transferring style from domain. Additionally, after transfer pre-train model; enabled adapt more transformer were then employed generate soft pseudo-labels hard (via iterative training) update parameters through distillation learning. Thus, retained learned features while adapting Experimental results indicated that maps method Market-to-Duke, Duke-to-Market, Market-to-MSMT, Duke-to-MSMT domains 70.2, 79.3, 30.2, 33.4, respectively.