Real-World Person Re-Identification via Super-Resolution and Semi-Supervised Methods
Limin Xia, Jiahui Zhu, Zhimin Yu · IEEE Access · 2021
Person re-identification has made great progress over the years. However, due to the problem of super-resolution and few labeled samples, it is difficult to apply in practice. In this paper, we propose a semi-supervised super-resolution person re-identification method based on soft multi-labels. Firstly, a Mixed-Space Super-Resolution model (MSSR) is constructed based on Generative Adversarial Networks (GAN), which aims to convert low-resolution person images into high-resolution images. Secondly, a Part-based Graph Convolutional Network (PGCN) is proposed to extract discriminative feature by exploring the relationship of local features within person. Finally, to solve the problem of label limitation, we use the PGCN trained with a small amount of labeled samples to predict the soft multi-labels of unlabeled samples, and further train PGCN with unlabeled samples based on a novel multi-label similarity loss. Experiments have been conducted on the Market1501, CUHK03, and MSMT17 datasets to evaluate this method, which show that it outperforms other semi-supervised methods.