论文名字:基于自适应对齐的可见红外多模态行人重识别

论文链接: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8913562.

1 摘要

可见光红外行人再识别(VI-Reid)除了由于摄像光谱不同造成的跨模态差异外,仍然存在较大的行人错位问题,以及传统的人再识别一样,由于摄像机视角的不同和行人姿态的不同而引起的变化,使得可见光红外人再识别(VI-Reid)仍然存在较大的不对准问题。
本文提出了一种 多路径自适应行人对齐网络(MAPAN) 来学习区分性特征表示。 端到端的方式并且不需要额外的手动注释即可自动对齐行人
为了减轻未对齐导致模态内的差异,联合了对齐过的可见光行人特征和原始的可见光行人特征来加强对行人的注意。
为了减轻可见光和红外交叉领域之间的差异,将两个模态的特征映射到相同的特征嵌入空间中。(上一篇多模态也是同样的处理手法)
损失:ID_LOSS+Triplet_Loss

2 理论研究(MAPAN)

Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment

2 结果分析

Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment
Visible Infrared Cross-Modality Person Re-Identification Network Based on Adaptive Person Alignment

相关文章:

  • 2021-10-12
  • 2021-04-12
  • 2021-03-31
  • 2021-09-21
  • 2021-11-28
  • 2021-12-28
  • 2021-04-23
  • 2021-04-16
猜你喜欢
  • 2021-05-31
  • 2020-04-06
  • 2021-06-16
  • 2021-06-11
  • 2021-10-14
  • 2021-11-16
  • 2021-09-03
相关资源
相似解决方案