基于域分离和对抗学习的跨域行人重识别

Cross⁃domain person re⁃identification based on domain separation and adversarial learning

针对跨域行人重识别应用中源域与目标域差异较大、现有模型无法在剥离域信息的同时有效获取关键身份信息的问题,提出一种基于对抗学习分离图像域信息与身份信息的方法.该方法由域分离和对抗学习两个阶段构成:域分离阶段分离图像行人特征和域特征;对抗学习阶段通过特征提取器与相机分类器的对抗学习,提升模型对域信息与身份信息的区分能力.在Market⁃1501,DukeMTMC⁃reID和MSMT17数据集上开展跨域行人重识别验证实验,实验结果表明,所提方法在跨域行人重识别任务上取得了显著的性能提升.

In the application of cross domain person re⁃identification,the source domain and the target domain are quite different,and the existing models can not effectively obtain the key identity information while stripping the domain information.This paper proposes a method to separate the image domain information and identity information based on adversarial learning.The method consists of two phases:domain separation and adversarial learning.Domain separation phase separates pedestrian features and domain features of image.In the stage of adversarial learning,feature extractor and camera classifier are used to improve the ability of distinguishing domain information from identity information.Cross domain person re⁃identification experiments are carried out on Market⁃1501,DukeMTMC⁃reID and MSMT17 datasets.The experimental results show that the proposed method achieves significant performance improvement on cross domain person re⁃identification task.