2020-07-01 09:43:09

Paperhttps://arxiv.org/abs/2002.02295 

Datasethttps://drive.google.com/file/d/1yTYawRm4ap3M-j0PjLQJ--xmZHseFDLz/view?usp=sharing 

 

1. Motivation

本文定义了一个新的 reid 任务:“Cross-clothes Person  reID” problem。不同于常规基于 color的方法来做 reid,本文尝试从行人轮廓的角度出发来做。因为同一个人的形状,短期内可能不是变化很大,如下图所示:

Person Re-identification by Contour Sketch under Moderate Clothing Change 

 

 

但是由于轮廓信息并不是所有的模式都有用。所以,作者提出一种可学习的极坐标转换网络来自动选择和采样相对可靠的模式。此外,作者也设计了一种 angle specific extractor(ASE)模块来建模不同 channel 之间的关系,来学习更加细粒度的 angle-specific features。 

 

由于缺乏大型的数据集,作者提出一个数据集,包含 221 个行人,33698 张图。

 

2. Approach Details

Person Re-identification by Contour Sketch under Moderate Clothing Change 

 

Person Re-identification by Contour Sketch under Moderate Clothing Change 

 

Person Re-identification by Contour Sketch under Moderate Clothing Change 

 

Person Re-identification by Contour Sketch under Moderate Clothing Change

 

 Person Re-identification by Contour Sketch under Moderate Clothing Change 

 

Person Re-identification by Contour Sketch under Moderate Clothing Change

Person Re-identification by Contour Sketch under Moderate Clothing Change

 

   

相关文章:

  • 2021-11-28
  • 2021-06-05
  • 2021-07-18
  • 2021-08-01
  • 2022-12-23
  • 2021-05-06
  • 2021-08-27
  • 2021-04-07
猜你喜欢
  • 2021-08-25
  • 2021-07-29
  • 2022-12-23
  • 2021-06-30
  • 2021-05-07
  • 2022-12-23
相关资源
相似解决方案