基于关系挖掘的跨模态行人重识别

A Cross-Modal Person Re-Identification Based on Relationship Mining

基于文本的行人重识别模型通常依赖于全局特征对齐和局部特征对齐,但模态间和模态内的相关信息常被忽略。提出了一种基于关系挖掘的跨模态行人重识别方法,该方法包括双流主干网络、负相似度挖掘模块、关系编码器。首先,通过双流主干网络实现了全局和局部特征对齐;其次,通过负相似度挖掘模块提升了图像-文本对特征辨别的细粒度;最后,通过关系编码器模块分别学习图像和文本中隐含的关系信息,实现关系级别的特征对齐。在CUHK-PEDES数据集和ICFG-PEDES数据集上的实验结果证明,文中方法能够达到较高的识别精度。

Aimed at the problems that while paying attention to the text-based person re-identification models often relying on global and local feature in alignment,very often the inter--modal and intra-modal correlations are in negative,a cross-modal pedestrian re-identification method is proposed based on relationship mining.The method includes a dual-stream network backbone,negative similarity mining module,and relationship encoder module.Firstly,the global and the local feature are in alignment through the dual-stream network backbone.Secondly the granularity of feature discrimination is enhanced by using the negative similarity mining module,and the similar incorrect results are filtered out.Finally,the relationship encoder module is utilized for respectively learning the latent relationship information in both the image and text,achieving relationship-level feature alignment.The experimental results on the CUHK-PEDES dataset and the ICFG-PEDES dataset show that this method achieves recognition accuracy higher.