Liu Haijun ; Tan Xiaoheng ; Zhou Xichuan

Abstract

This paper focuses on the visible-thermal cross-modality personre-identification (VT Re-ID) task, whose goal is to match person images betweenthe daytime visible modality and the nighttime thermal modality. The two-streamnetwork is usually adopted to address the cross-modality discrepancy, the mostchallenging problem for VT Re-ID, by learning the multi-modality personfeatures. In this paper, we explore how many parameters of two-stream networkshould share, which is still not well investigated in the existing literature.By well splitting the ResNet50 model to construct the modality-specific featureextracting network and modality-sharing feature embedding network, weexperimentally demonstrate the effect of parameters sharing of two-streamnetwork for VT Re-ID. Moreover, in the framework of part-level person featurelearning, we propose the hetero-center based triplet loss to relax the strictconstraint of traditional triplet loss through replacing the comparison ofanchor to all the other samples by anchor center to all the other centers. Withthe extremely simple means, the proposed method can significantly improve theVT Re-ID performance. The experimental results on two datasets show that ourproposed method distinctly outperforms the state-of-the-art methods by largemargins, especially on RegDB dataset achieving superior performance,rank1/mAP/mINP 91.05%/83.28%/68.84%. It can be a new baseline for VT Re-ID,with a simple but effective strategy.