论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification

时间:2023-12-16 21:41:38

论文阅读笔记(十七)ICCV2017的扩刊(会议论文【传送门】)

改进部分:

(1)惩罚函数:原本由两部分组成的惩罚函数,改为只包含 Sequence Cost 函数;

(2)对重新权重改进:

① Positive Re-Weighting:

论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification

其中 论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification若太大,则选择的样本标签的可信度小;论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification若太小,则样本数量不足以进行矩阵学习,因此设置如下的论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification

论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification

其中,σ为 [0, 1],如果 σ = 1,则说明充分相信样本估计的可信度,反之设置为 σ = 0.

② Negative Re-Weighting:

对于所有被估计为负的样本对,全部设置为:

论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification

(后续进行矩阵学习过程不变)

(3)Co-Matching策略:

将视频序列划分为两个子序列,收敛为两个聚类,标签设置如下:

论文阅读笔记(十八)【ITIP2019】:Dynamic Graph Co-Matching for Unsupervised Video-Based Person Re-Identification