基于多度量融合的无监督领域自适应行人重识别算法  

Unsupervised domain adaptive person Re-ID algorithm based on multi-metric fusion

作　　者：姜冠正 唐俊[1] JIANG Guanzheng;TANG Jun(School of Electronic Information Engineering,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学电子信息工程学院,安徽合肥230601

出　　处：《安徽大学学报（自然科学版）》2023年第1期29-38,共10页Journal of Anhui University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61772032);安徽省重点研究与开发计划项目(202004a07020050)。

摘　　要：行人重识别任务旨在跨相机下检索出特定的行人图像.虽然行人重识别任务得到了快速发展,在检索精度上得到很大的提升,但是依然面临着行人重识别模型在新的数据集上泛化能力有限,以及在无监督领域自适应任务中无法避免的伪标签噪声的问题.针对目前无监督领域自适应任务中由于聚类算法的局限性而导致伪标签出现噪声的问题,提出一种基于多度量融合的无监督领域自适应行人重识别算法.具体而言,多度量融合算法是在目标域上使用DBSCAN(density-based spatial clustering of applications with noise)聚类算法对特征空间的行人特征进行聚类时,通过多个特征相似度度量函数线性加权的方式,计算行人之间的特征相似度,从而在目标域上生成更为准确的伪标签,之后利用该伪标签微调模型.通过在Market1501→DukeMTMC-reID和DukeMTMC-reID→Market1501上大量的实验,证明多度量融合算法有效提升了行人重识别模型在无监督领域自适应任务上的检索精度.The person re-identification task aims to retrieve images of specific pedestrians across cameras.Although the person re-identification task has developed rapidly and the retrieval accuracy has been greatly improved,the person Re-ID model still faces the problem of limited generalization ability on new data sets and unavoidable pseudo label noise in unsupervised field adaptive tasks.Aiming at the problem of noise in pseudo-labels due to the limitations of clustering algorithms in current unsupervised domain adaptation tasks,this paper proposes an unsupervised domain adaptive person re-identification algorithm based on multi-metric fusion.Specifically,the multi-metric fusion algorithm uses the DBSCAN(density-based spatial clustering of applications with noise)clustering algorithm on the target domain to cluster pedestrian features in the feature space and calculates the feature similarity between pedestrians by linearly weighting multiple feature similarity metric functions,so as to generate more accurate pseudo-labels on the target domain,and then uses the pseudo-labels to fine-tune the model.Through a large number of experimental results on Market1501→DukeMTMC-reID and DukeMTMC-reID→Market-1501,it is shown that the multi-metric fusion algorithm effectively improves person re-identification model retrieval accuracy on unsupervised domain adaptation tasks.

关 键 词：多度量融合 伪标签 行人重识别 无监督领域自适应 特征相似度 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]