GAN-based data augmentation and pseudo-label refinement with holistic features for unsupervised domain adaptation person re-identification

Abstract

Person re-identification (re-ID) by using an unsupervised domain adaptation (UDA) approach has drawn considerable attention in contemporary security research. Thus, UDA person re-ID usually employs a model learned from a labeled source domain, adjusted by pseudo-labels, for an unlabeled target domain. However, it still needs to overcome three main challenges: the significant distribution gap between the source and target domain data, the disregarding of important information during the learning process, and the accuracy of pseudo-labels generated by a clustering algorithm. To tackle these problems, we propose a novel method to improve UDA person re-ID performance by combining GAN-based data augmentation and unsupervised pseudo-label refinement with holistic features method for training on target domain, named DAPRH. In particular, we first use a generative adversarial network (GAN) method to generate style-transferred images with the target domain style to enrich the training set and the domain invariant mapping (DIM) technique to effectively mitigate the distribution gap between the source and target domains. Next, we also utilize information from partial regions along with the global feature as holistic features to ensure that all necessary information is obtained during the learning process. Finally, we propose a centroid-based pseudo-label refinement method using a soft target label (one-soft) instead of the originally fixed label (one-hot) to enhance the teacher-student architecture-based unsupervised learning process. Extensive experiments on three well-known datasets, Market-1501, DukeMTMC-reID, and MSMT17, demonstrate that DAPRH can significantly surpass the state-of-the-art performance of the UDA person re-ID.