Visible Embraces Infrared: Cross-Modality Person Re-Identification with Single-Modality Supervision
Jiangming Shi, Xiangbo Yin, Demao Zhang, Yanyun Qu · 2023
Visible-infrared person re-identification (VI-ReID) has garnered significant attention due to its wide range of applications. However, a substantial performance gap still ex-ists between cross-modality and conventional ReID methods. Furthermore, existing VI-ReID models heavily rely on super-vised learning, necessitating a substantial amount of labeled infrared data. To overcome these limitations, we explore a “uni-to-cross” setting for VI-ReID, where the labels of infrared images are inaccessible. Our observation is that existing cross-modality datasets have a relatively small scale, while visible ReID datasets offer rich annotation. In this paper, we propose a disentangled generative solution for a “uni-to-cross” setting, which includes a disentangled representation learning module and a discriminative feature learning module. The former purifies the id-related information by utilizing a disentangled generative network. Meanwhile, the latter enables the network to learn from synthesized infrared images, effectively reducing the gap between the two modalities. Extensive experimental results on SYSU-MM01 and RegDB demonstrate that the proposed method gains impressive performance. Especially, our method achieves breakthrough improvement by a large margin of 10.8% and 10.4% in terms of mAP on SYSU-MM01 and RegDB datasets, respectively.