Deep learning based methods for low-resolution person re-identification

• Main Author: Sebastian, Jason
• Other Authors: Lin Weisi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/147938
• Institution: Nanyang Technological University

Person re-identification (re-id) is a fundamental task in automated surveillance. In practice, surveillance is often performed using multiple camera views to expand surveillance range. Hence the resolution of person images captured under these different camera views varies drastically due to the unconstrained distance between persons and cameras. We call this setting low-resolution person re-identification (LR re-id). In LR re-id, query images are often of low-resolution (LR) and gallery images are often of high-resolution (HR). This causes the resolution mismatch problem, which needs to be addressed because LR query images lack appearance details which are needed for re-id. It is a non-trivial problem because simply resizing the query image does not recover appearance details. Recently, HRNet was introduced as a deep learning framework to tackle position-sensitive tasks which require high-resolution representations. HRNet has shown state-of-the-art results in semantic segmentation, object detection, etc. Although the possibility of implementing HRNet for applications in the LR re-id setting looks promising, to the best of our knowledge there has yet to be an attempt to apply HRNet for LR re-id. Moreover, with advancements in super-resolution (SR) research, it is possible to recover appearance details from LR images which are needed to solve the resolution mismatch problem. This project attempts to develop an effective deep learning framework to cope with the LR re-id problem. This consists of three parts. Firstly, use HRNet as a feature extraction module for re-id. Secondly, use an existing deep learning based SR module to enhance appearance details of a given query image. Thirdly, jointly learn SR and re-id modules so that the appearance details generated by the SR module are useful for the re-id feature extraction task. This project successfully demonstrates that it is possible to address the resolution mismatch problem by using HRNet as a backbone and by jointly learning SR and re-id.