1.昆明理工大学 信息工程与自动化学院 昆明;2.昆明理工大学 信息工程与自动化学院
TP311
云南省科技厅科技计划项目(面上项目)(202101AT070136)和云南省基础研究计划项目（202301AV070004）
Faculty of Information Engineering and Automation,Kunming University of Science and Technology
Supported by Science and Technology Planning Project of Yunnan Science and Technology Department (General Project) (202101AT070136) and Supported by Basic Research Program Project of Yunnan Province (202301AV070004)
红外-可见光行人重识别（VI-ReID）是一项在可见光和红外光模态下匹配相同行人的任务。尽管现有的 VI-ReID 方法取得一定成果，但这些方法并未有效利用行人的跨模态显著信息和多粒度信息。为此，本文提出了模态显著一致性特征挖掘和多粒度特征增强的 VI-ReID 方法。其中，模态显著一致性特征挖掘改进了交叉注意力机制的关系权重选择方式，实现了模态显著一致性特征的有效提取；多粒度特征增强通过提取行人不同粒度下的细粒度特征，有效利用了行人的局部鉴别性信息。与现有方法相比，本文提出的方法在公共数据集 SYSU-MM01 和 RegDB 上取得了明显的优势。实验表明，跨模态显著信息和多粒度信息的有效利用能够提升特征的跨模态鲁棒性。
Visible-Infrared person re-identification (VI-ReID) is a task of matching person in both visible and infrared modes. Although existing VI-ReID methods have achieved certain results, they have not effectively utilized the cross modality significant information and multi-granularity information of person. Therefore, this paper proposes a VI-ReID method for modal significant consistency feature mining and multi-granularity feature enhancement. Among them, the modal significant consistency feature mining improves the relationship weight selection method of cross attention mechanism, achieving effective extraction of modal significant consistency features; multi-granularity feature enhancement effectively utilizes the local discriminative information of person by extracting fine-grained features of person at different granularity. Compared with existing methods, the method proposed in this paper has achieved significant advantages on the public datasets SYSU-MM01 and RegDB. Experiments have shown that the effective utilization of cross modality significant information and multi-granularity information can enhance the cross modal robustness of features.
