Visible & infrared person re-identification

Abstract

Person Re-Identification (ReID) aims to match persons in different images with viewpoint and illumination variation captured by different cameras. However, visible cameras cannot work well at night or low illumination areas. Considering advanced surveillance cameras automatically switch to infrared mode when detecting poor light conditions, we expanded the existing single-modality RGB-based visible Person ReID problem to the cross-modality (visible and infrared (IR) images) problem. In this dissertation, we research two existing large-scale Visible & Infrared Person Re-Identification (VI-ReID) datasets, existing methods based on different Convolutional Neural Networks (ConvNet), and compare their characteristics and performances. Besides, we introduce a new evaluation protocol and propose a new baseline for the research. This work demonstrates the existing methods are capable of dealing with the VI-ReID problem, but their performances still have room for improvement compared with the singlemodality RGB-based Person ReID problem. In addition, our new evaluation protocol uses IR images as queries and RGB images as the gallery set in the SYSU-MM01 dataset, which is more suitable for evaluating if the methods are biased. Moreover, Residual Neural Network (ResNet) based methods, including our baseline, are better than Generative Adversarial Network (GAN) based methods in both accuracy and efficiency, which shed light on research into new network structures in the future.