A Knowledge-Driven Enhanced Module for Visible-Infrared Person Re-identification

Shihao Shan, Enyuan Xiong, Xiang Yuan, Song Wu

Abstract: Visible-infrared person re-identification (VI-ReID) is a task in computer vision that has gained increasing importance in today’s social surveillance systems. VI-ReID suffers from additional cross-modality variance caused by the inherent heterogeneous gap between the visible and infrared modality compared with previous person re-identification. This paper aims to reduce the difficulty of cross-modality shared feature learning through a knowledge-driven modality. From a novel perspective, we focus on explicitly modeling an appropriate knowledge-driven modality to narrow the cross-modality difference. Specifically, we propose a Knowledge-driven Enhanced Module (KDEM) to synthesize the feature of the knowledge-driven modality and help the model accumulate transitional knowledge. The synthetic feature distribution is controlled by two modality influencing factors generated by KDEM. To this end, for better characterizing knowledge-driven modality in a diverse way, we enforce a diversity loss on the two modality influence factors, which is instrumental in the knowledge accumulation of the model. Meanwhile, the Consistency loss is proposed to maintain the similarity between the knowledge-driven modality and the other two modalities, thereby can avoid knowledge accumulation of the knowledge-driven modality impacts on the performance of the model. In all the common VI-REID tasks, Our proposed method performs state-of-the-art. The code will be announced at https://github.com/SWU-CS-MediaLab/KDEM.

External IDs:dblp:conf/icann/ShanXYW22