Parameter Sharing Exploration and Hetero-Center Triplet Loss for Visible-Thermal Person Re-Identification

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