A Comprehensive Survey and Outlook for Cross-Resolution Person Re-Identification

Abstract

Person re-identification (Re-ID) is a fundamental task in computer vision which has achieved significant progress in recent years. However, the existing promising algorithms are typically based on the assumption that all the images have the same and sufficiently high resolution (HR), ignoring the fact that the images are often captured with different resolutions. This study intends to present a comprehensive overview of cross-resolution (CR) person Re-ID to promote a deeper understanding of this topic and further research. We first group the current techniques into three categories: dictionary-learning-based, super-resolution-based, and generative-adversarial-network-based methods. The motivation, principles, benefits, and drawbacks of these techniques are extensively discussed. Then, the ways to construct synthetic multi-low-resolution (MLR) datasets and the performance comparisons of the state-of-the-art algorithms on five MLR datasets are demonstrated. Finally, challenges and potential research directions are further discussed.