基于图卷积神经网络的跨域行人再识别  

Cross-domain person re-identification using graph convolutional networks

作　　者：潘少明[1] 王玉杰 种衍文[1] PAN Shaoming;WANG Yujie;CHONG Yanwen(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079

出　　处：《华中科技大学学报（自然科学版）》2020年第9期44-49,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(41671382,61572372,41271398);国家重点研发计划资助项目(2017YFB0504202)。

摘　　要：针对由源域训练的行人再识别模型通常在目标域的泛化能力不强的问题,提出基于图卷积神经网络的跨域行人再识别方法,将源域数据学习到的整合邻居样本信息的能力迁移至目标域数据.首先,为经过特征提取后的源域数据建立亲属子图,并将源域数据特征和亲属子图作为所设计的图卷积神经网络模块的输入,以基于源域的监督信息训练图卷积神经网络模块;然后,对经过特征提取后的目标域数据建立亲属子图,将训练过的图卷积神经网络模块应用于目标域数据,为目标域数据赋伪标签;最后,联合源域数据和目标域数据训练得到一个泛化能力强的行人再识别模型.分别在两个大规模公开数据集Market-1501和DukeMTMC-reID上对所提出方法进行实验验证,结果表明所提出的方法与所选择的基准模型相比使得Market-1501的rank-1准确率和平均准确率均值(mAP)分别提高了7.4%和9.2%,而DukeMTMC-reID的rank-1准确率和m AP分别提高了14.2%和14.9%.Considering the problem that the person re-identification model trained by the source domain usually can only obtain weak generalization ability in target domain,a method based on graph convolutional networks(GCN)was proposed for cross-domain person re-identification by transferring the ability of integrating neighbor sample information learned from the source domain to target domain.Firstly,a source affinity subgraph was established based on data features of source domain.Then the source affinity subgraph and data features of source domain were taken together as the input of the designed graph convolutional neural network module so as to train the module based on the supervisory information of the source domain.Secondly,after establishing the target affinity subgraph based on data features of target domain,the target affinity subgraph and the trained graph convolution neural network module can be used to realize the purpose of assigning pseudo labels for the target domain data.Lastly,a generalized person re-identification model can be obtained by combining the source domain data and the target domain data.Experiments are constructed on two large public dataset:Market-1501 and DukeMTMC-reID.As shown in the extensive experimental results,compared with the baseline model,the rank-1 accuracy and mean average precision(mAP)on Market-1501 is improved 7.4%and 9.2%,respectively;the rank-1 accuracy and mAP on DukeMTMC-reID is improved 14.2%and 14.9%,respectively.

关 键 词：行人再识别 跨域 图卷积神经网络 亲属图 深度学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]