Cross-modality person re-identification via channel-based partition network

摘要

Visible-infrared cross-modality person re-identification is an important task in the night video surveillance system, the huge difference between infrared and visible light images makes this work quite challenging. Unlike traditional person re-identification, a cross-modality mission needs to solve intra-class differences and inter-class variations. To solve the problem of huge modality discrepancy, in this paper, we propose a channel-based partition network which can unify the features of the two modes in an end-to-end way. Firstly, to handle the lack of discriminative information, we introduce newly generated samples to help the network improve its ability to learn cross modal features. Secondly, at the feature level, we propose a distinctive method of learning local features, in which the set of feature maps is parted on the channel. At the end of the proposed framework, we add a lightweight feature converter to further eliminate modality differences. The experimental results on the two popular datasets prove the effectiveness of our work.