基于相机感知多样代理的无监督行人重识别  

Camera-aware Diverse Proxies Based Unsupervised Person Re-identification

作　　者：郁航 范慧杰[2] 付生鹏[2] 陈希爱 赵立杰 YU Hang;FAN Huijie;FU Shengpeng;CHEN Xi’ai;ZHAO Lijie(School of Information Engineering,Shenyang University of Chemical Technology,Shenyang 110142,China;State Key Laboratory of Robotics,Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China)

机构地区：[1]沈阳化工大学信息工程学院,辽宁沈阳110142 [2]中国科学院沈阳自动化研究所机器人学国家重点实验室,辽宁沈阳110016

出　　处：《机器人》2025年第2期169-178,共10页Robot

基　　金：国家自然科学基金(62273339,62073205);辽宁省重点研发计划(2024JH2/102400022);中国科学院基础研究项目.

摘　　要：基于聚类的无监督行人重识别通过聚类产生伪标签,并以此为每个聚类构建单一代理。但不同相机之间的差异以及同一相机下行人不同姿态方向等导致的相机类内差异使得单一代理无法充分表示整个复杂聚类,进而影响训练过程和模型的性能。为此,本文提出基于相机感知多样代理的无监督行人重识别方法来消除相机差异和相机类内差异,以提高类内样本的紧密性。相机感知多样代理包括相机多代理和相机差异代理。其中相机多代理为每个相机类构建一个平均代理和一个硬代理,通过互补表示来提高相机类内样本的紧密性。同时还会为每个相机类构建一个相机差异代理,相机差异代理会利用不同相机类的平均代理进行更新,以此拉近相机之间的距离,减少相机差异。相机差异代理和相机多代理相互协作,提高类内样本紧密性和模型性能。本文方法在Martket1501、DukeMTMC-reID、MSMT17三个数据集上mAP/Rank-1分别达到了90.4%/96.4%、75.8%/86.5%、58.2%/82.5%。Cluster-based unsupervised person re-identification generates pseudo labels through clustering and constructs a single proxy for each cluster.However,inter-camera discrepancies between different cameras and intra-camera differences caused by pedestrians’different stepping directions under the same camera make it impossible for a single proxy to adequately represent the whole complex cluster,affecting the training process and model performance.Thus,a camera-aware diverse proxies based unsupervised person re-identification method is proposed to eliminate inter-camera discrepancies and intracamera differences in order to improve the intra-class tightness.Camera-aware diverse proxies include camera multi proxies and camera discrepancy proxy.The camera multi proxies construct a mean proxy and a hard proxy for each camera class to improve intra-camera tightness through complementary representation.A camera discrepancy proxy is also constructed for each camera class,which utilizes the mean proxies of different camera classes for updating,so as to bring cameras closer to each other and reduce camera discrepancy.The camera discrepancy proxy and camera multi proxies collaborate with each other to improve intra-class tightness and model performance.The proposed method achieves 90.4%/96.4%,75.8%/86.5%,and 58.2%/82.5%for mAP/Rank-1 on the Martket1501,DukeMTMC-reID,and MSMT17 datasets,respectively.

关 键 词：行人重识别 无监督 相机差异代理 相机多代理 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]