Orthogonal Deep Feature Decomposition Network for Cross-Resolution Person Re-Identification

Abstract

Person images captured by surveillance cameras in real scenes often have low resolution (LR), which suffers from severe degradation in recognition performance when matched with pre-stocked high-resolution (HR) images. There are existing methods which typically employ super-resolution (SR) techniques to address the resolution discrepancy problem in person re-identification (re-ID). However, SR techniques are intended to enhance the human eye visual fidelity of images without caring about the recovery of pedestrian identity information. To cope with this challenge, we propose an orthogonal depth feature decomposition network. And we decompose pedestrian features into resolution-related features and identity-related features who are orthogonal to each other, from which we design the identity-preserving loss and resolution-invariant loss to ensure the recovery of pedestrian identity information. When compared with the SOTA method, experiments on the MLR-CUHK03 and MLR-VIPeR datasets demonstrate the superiority of our method.