首页|基于相关性得分的伪标签优化行人重识别

基于相关性得分的伪标签优化行人重识别

扫码查看
无监督域自适应行人重识别旨在将源域训练的识别能力泛化到目标域上,以减少对标签的依赖.目前基于聚类方法的网络,聚类过程中不可避免地会受到环境噪声的影响,降低网络原有识别性能.为了解决这一问题,提出一种基于相关性得分的伪标签优化行人重识别网络.首先,通过计算全局与局部特征间前k个相似样本集合的相关性得分,找到两类特征直接可靠的关联性,从而提取已有伪标签优化方法所忽略的局部细粒度特征;然后,利用得分对局部伪标签进行优化处理,降低网络对与行人无关局部特征的关注;最后,依赖于相关性得分,利用优化后局部伪标签的预测结果对全局伪标签进行细化,缓解聚类过程中噪声的同时也细化了行人的特征完整表示.与近年无监督域自适应行人重识别方法相比,所提出网络在DukeMTMC-ReID、Market 1 501和MSMT17三个公开数据集上的实验结果表现优异,验证了所提出网络的有效性.
Person re-identification with pseudo label refinement based on correlation score
The purpose of unsupervised domain adaptive person re-identification is to generalize the recognition ability of training in the source domain to the target domain to reduce the dependence on labels.At present,the network based on clustering methods will inevitably be affected by environmental noise during the clustering process,which will reduce the original recognition performance of the network.To solve this problem,the person re-identification network with pseudo label refinement based on correlation score is proposed.Firstly,by calculating the correlation scores between the top k similar sample sets of global and local features,reliable correlations between two types of features are found,so as to extract local fine-grained features that existing pseudo-label optimization methods ignore.Then,the scores are used to optimize the local pseudo-labels,reducing the network's attention to irrelevant local features of the person.In addition,relying on the correlation scores,the prediction results of the optimized local pseudo-labels are used to refine the global pseudo-labels,which alleviates noise during the clus-tering process and refines the complete representation of person features.Compared with the unsupervised domain adaptive method in recent years,the experimental results of the network on three public data sets,DukeMTMC-ReID,Market 1 501 and MSMT17,show that the network performance is significantly improved.

unsupervised domain adaptationperson re-identificationdeep learningfine-grained featurescorrelation scorelocal pseudo label refinement

程德强、黄绩、寇旗旗、张剑英、李云龙

展开 >

中国矿业大学信息与控制工程学院,江苏徐州 221116

中国矿业大学计算机科学与技术学院,江苏徐州 221116

无监督域自适应 行人重识别 深度学习 细粒度特征 相关性得分 伪标签优化

国家自然科学基金项目

52204177

2024

控制与决策
东北大学

控制与决策

CSTPCD北大核心
影响因子:1.227
ISSN:1001-0920
年,卷(期):2024.39(8)