知识引导和细粒度信息增强的无监督域自适应行人再识别
DOI:
作者:
作者单位:

1.昆明理工大学信息工程与自动化学院;2.昆明理工大学 信息工程与自动化学院

作者简介:

通讯作者:

中图分类号:

TP311

基金项目:

国家自然科学基金(61966021,61562053),大学生创新创业训练计划项目(202010674098)


Knowledge guidance and fine-grained information enhancement for Unsupervised domain adaptation person re-identification
Author:
Affiliation:

Faculty of Information Engineering and Automation,Kunming University of Science and Technology

Fund Project:

National Natural Science Foundation of China(61966021,61562053),College Students' Innovative Entrepreneurial Training Plan Program(202010674098)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    无监督域自适应旨在将源域知识迁移到目标域数据集上,是行人再识别中一个非常重要的任务。现实情况下,采集到的数据往往具有视频帧信息。因此,目标域数据集中单一视角下的行人标签极易获取。然而,已有的方法忽略了这些信息,从而限制了识别性能的提升。为此,提出知识引导和细粒度信息增强的无监督域自适应行人再识别。在该方法中,开发了目标域单视角下行人样本已知的新范式,充分挖掘了目标域中有用的信息。同时,将源域知识作为引导,辅助模型提取目标域行人图像的判别性特征。与已有的知识迁移策略相比,这种知识引导的方式能更加有效地缓解域偏移。进一步,将行人的局部信息嵌入到全局特征中,增强了特征的细粒度信息。在两个公开的数据集上进行实验,充分证明了提出方法的有效性和优越性。代码地址:https://github.com/lhf12278/KG-FGIN

    Abstract:

    Unsupervised domain adaptation aims to transfer the source domain knowledge to the target domain dataset, which is a very important task in person re-identification. In reality, the collected data often has video frame information. Therefore, it is very easy to obtain pedestrian labels from a single camera view in the target domain dataset. However, the existing methods ignore the available information, which limits the improvement of recognition performance. For this reason, knowledge guidance and fine-grained information enhancement for unsupervised domain adaptation person re-identification is proposed. In this method, a new paradigm with a single-view pedestrian sample known in the target domain is developed to fully mine the useful information in the target domain. At the same time, the source domain knowledge is used as a guide to assist the model to extract the discriminative features of pedestrian images in the target domain. Compared with the existing knowledge transfer strategies, this knowledge-guided approach can effectively alleviate the domain shift. Furthermore, the local information of pedestrians is embedded into the global feature, and the fine-grained information of the feature is enhanced. Experiments on two publicly datasets fully prove the effectiveness and superiority of the proposed method. Code: https://github.com/lhf12278/KG-FGIN

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-04-19
  • 最后修改日期:2021-05-08
  • 录用日期:2021-05-25
  • 在线发布日期:
  • 出版日期: