1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记

作者文章参考自:

S.-Z. Chen, C.-C. Guo, and J.-H. Lai, “Deep ranking for person re-identifcation via joint representation learning,” IEEETransactions on Image Processing, vol. 25, no. 5, pp. 2353–2367,2016.其顶层采用的是ranking layer,作者采用了其样本对作为一张图片输入的网络架构
Y. Tang, Deep Learning Using Support Vector Machines, CoRR,2013 参考了向量机部分,引入了可微的L2-SVM,并用一个分支代替了两个分支输入。

网络架构如下:
1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记_第1张图片
1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记_第2张图片

实验在CUHK01和VIPeR上进行,实验效果一般!

训练用的一些技巧:DropOut,BN,Data Augmentation 和初期训练的样本对 Data Balancing,预训练在CUHK02上,在CUHK01和VIPeR上测试时会finetune 预训练模型的最后几层。
1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记_第3张图片

其他实验验证:
Superiority of Joint Representation Learning
1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记_第4张图片

Superiority of Linear SVM Layer
1707.Deep Learning for Person Reidentification Using Support Vector Machines 论文笔记_第5张图片

总结:
文中关于网络层公式的推导和梯度反传推导值得参考。

你可能感兴趣的:(行人重识别)