[论文笔记]CVPR2016_Person re-identifcation by multi-channel parts-based cnn with improved triplet loss f

Title: Person re-identifcation by multi-channel parts-based cnn with improved triplet loss function

Authors: De Cheng, Yihong Gong, Sanping Zhou, Jinjun Wang, Nanning Zheng

Affiliations: Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University,Xi’an, Shaanxi, P.R. China


Contribution

  1. 好像之前的文章都是孪生网络,这篇文章第一次引入了triplet framework,设计了一种网络使得输入三张图像而非两张
  2. 对人脸识别领域的triplet loss function进行改进,使得更适用于person reid问题
  3. 在多个数据集上都达到了state-of-the-art performance

具体方法

网络结构

网络输入不再是两张图像,而是三张图像。对于第i个输入的图像组,记做$I_i=<I_{i}^{o},I_{i}^{+},I_{i}^{-}>$。其中,$I_{i}^{o}$和I_{i}^{+}是同一个人,而$I_{i}^{-}$是不同的人。

[论文笔记]CVPR2016_Person re-identifcation by multi-channel parts-based cnn with improved triplet loss f

 

上一篇:Problem 9


下一篇:LC.334. Increasing Triplet Subsequence