首页|基于动态辅助对比学习的跨域行人重识别

基于动态辅助对比学习的跨域行人重识别

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
具有混合记忆的自步对比学习(Self-paced Contrastive Learning,SpCL)通过集群聚类生成不同级别的伪标签来训练网络,取得了较好的识别效果,然而该方法从源域和目标域中捕获的行人数据之间存在典型的分布差异,使得训练出的网络不能准确区别目标域和源域数据域特征.针对此问题,提出了双分支动态辅助对比学习(Dynamic Auxiliary Contrastive Learning,DACL)框架.该方法首先通过动态减小源域和目标域之间的局部最大平均差异(Local Maximum Mean Discrepancy,LMMD),以有效地学习目标域的域不变特征;其次,引入广义均值(Generalized Mean,GeM)池化策略,在特征提取后再进行特征聚合,使提出的网络能够自适应地聚合图像的重要特征;最后,在3 个经典行人重识别数据集上进行了仿真实验,提出的DACL与性能次之的无监督域自适应行人重识别方法相比,mAP和rank-1 在Market1501 数据集上分别增加了 6.0 个百分点和 2.2 个百分点,在MSMT17 数据集上分别增加了2.8 个百分点和3.6 个百分点,在Duke数据集上分别增加了 1.7 个百分点和2.1 个百分点.
Dynamic auxiliary contrastive learning for cross-domain person re-identification
The self-paced contrastive learning(SpCL)with hybrid memory uses clustering to train the network to generate different levels of pseudo labels,and achieves good re-identification results.However,as a distribution difference exists between the pedestrian datasets captured from the source domain and the target domain,the trained network cannot accurately identify the features of the target domain and the source domain.To solve this problem,this paper proposes a two-branch dynamic auxiliary contrastive learning(DACL)framework.This framework effectively learns the domain invariant features of the target domain by dynamically reducing the local maximum mean discrepancy(LMMD)between the source domain and the target domain.In addition,the generalized mean(GeM)pooling strategy is adopted to aggregate features after feature extraction,so that the proposed network can adaptively aggregate the important features in the image space.Finally,simulation experiments are conducted on three classic person re-identification datasets.Compared with the unsupervised domain adaptation for person re-identification methods with the second best performance,the proposed DACL increases mAP and rank-1 by 6.0 percentages and 2.2 percentages on the Market1501 dataset,2.8 percentages and 3.6 percentages on the MSMT17 dataset,and 1.7 percentages and 2.1 percentages on the Duke dataset,respectively.

person re-identification(ReID)unsupervised domain adaptation(UDA)generalized mean(Gem)poolinglocal maximum mean discrepancy(LMMD)contrastive learning

杨真真、邵静、杨永鹏、吴心怡

展开 >

南京信息职业技术学院网络与通信学院,江苏 南京 210023

南京邮电大学 宽带无线通信与传感网技术教育部重点实验室,江苏 南京 210003

行人重识别 无监督域自适应 广义均值池化 局部最大平均差异 对比学习

国家自然科学基金国家自然科学基金江苏省研究生科研与实践创新计划项目江苏省研究生科研与实践创新计划项目南京邮电大学科研项目

6207124262171232KYCX22_0955SJCX23_0251NY220207

2024

10.14132/j.cnki.1673-5439.2024.03.008

南京邮电大学学报(自然科学版)
南京邮电大学

南京邮电大学学报(自然科学版)

CSTPCD北大核心
影响因子：0.486
ISSN：1673-5439
年,卷(期)：2024.44(3)
  • 28