基于纯无监督的行人重识别方法研究

Research on the Person Re-identification Methods Based on Purely Unsupervised

行人重识别技术在智慧安防领域占据重要的地位。其中,领域自适应行人重识别依赖于有监督的源域进行预训练,而纯无监督行人重识别则完全不依赖行人身份标签。文章分析基于轨迹和基于聚类的纯无监督行人重识别方法,并探讨近年来在伪标签处理上的重要进展。通过对CUHK03、Market-1501、DukeMTMC-Re-ID、MSMT17等常用数据集的介绍分析,总结当前纯无监督行人重识别所面临的挑战与未来研究方向,包括自监督学习、解决模型训练中的灾难性遗忘问题和构建高质量数据集等。

Person re-identification(Re-ID)technology is crucial in the field of intelligent security and protection.Domain-adaptive person re-identification relies on supervised source domain pre-training,while purely unsupervised person re-identification completely eschews person identity labels.This paper provides a comprehensive overview of tracklet-based and clustering-based purely unsupervised person re-identification methods and explores recent advancements in pseudo-label processing.By introducting and analyzing widely used datasets such as CUHK03,Market-1501,DukeMTMC-ReID,and MSMT17,this paper summarizes the challenges and future research directions in purely unsupervised person re-identification,including self-supervised learning,addressing catastrophic forgetting in model training,and constructing high-quality datasets.