Visible-thermal person re-identification via multiple center-based constraints

Person re-identification is an important part of the intelligent video analysis and processing system. A new type of surveillance camera can switch to the thermal infrared mode for 24-hour video surveillance. It is necessary to research the visible-thermal cross-modality person re-identification. However, there is a large modal discrepancy in the visible-thermal task. Therefore, the research focus on how to build a bridge to narrow the cross-modality gap and fully exploit the shared information. In this paper, we first employ grayscale transformation of the visible image to generate an intermediate modality, thereby reducing the distance between the original two domains. On this basis, a novel three-branch multiple center-constrained network (TMCC-Net) is built for the visible-thermal task. More specifically, TMCC-Net is a three-branch network that mines the shared information of pedestrians in three modalities through special feature learning and shared feature embedding. In order to obtain better performance, our work introduces two heterogeneous center-constrained losses to constrain the feature embedding. On the one hand, the proposed losses limit the distribution of features at the modality edge; on the other hand, they can strengthen the role of grayscale modality in the cross-modality task. Finally, an end-to-end model for visible-thermal person re-identification is built, which is effective for shared information mining. Extensive experiments are conducted on the two cross-modality datasets, including the SYSU-MM01 and RegDB datasets. The experimental results demonstrate the effectiveness and superiority of the proposed method compared to the state-of-the-art approaches.