Unsupervised Cross-Domain Adaptation through Mutual Mean Learning and GANs for Person Re-identification

Unsupervised cross-domain adaptation is a challenging task for person re-identification due to the unavailability of target domain labels. Among existing methods, pseudo-Iabels-based methods have considerable performance but most of them use target domain data without labels which are challenging difficult for the target model to learn enough features. In this paper, we use generative based models that generate more target data. In cooperation with the generative model, a mutual learning model is used to transfer knowledge of one model to another model that ultimately improves overall model performance. Ex-tensive experiments are performed on Duke and Market datasets that significantly achieve improved performance in comparison to state-of-the-art methods.