Deep learning for visible-infrared cross-modality person re-identification: A comprehensive review

Visible-infrared cross-modality person re-identification (VI-ReID) is currently a prevalent but challenging research topic in computer vision, since it can remedy the poor performance of existing single-modality ReID models under insufficient illumination, thus enabling the 24/7 surveillance systems. Although extensive research efforts have been dedicated to VI-ReID, a systematic and comprehensive literature review is still missing. Considering that, in this paper, a comprehensive review of VI-ReID approaches is provided. First, we clarify the importance, definition and challenges of VI-ReID. Secondly and most importantly, we elaborately analyze the motivations and the methodologies of existing VI-ReID methods. Accordingly, we will provide a comprehensive taxonomy, including 4 categories with 8 sub-items, for those state-of-the-art (SOTA) VI-ReID models. After that, we elaborate on some widely used datasets and evaluation metrics. Next, comprehensive comparisons of SOTA methods are made on the benchmark datasets. Based on the results, we point out the limitations of current methods. At last, we outline the challenges in this field and future research trends.