Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录

基于深度学习的空间对齐方法在行人重识别中的应用

Digital Object Identifier 10.1 109/ACCESS.2019.2945353

论文链接: https://ieeexplore.ieee.org/document/8856191.
STN模块链接: https://github.com/kevinzakka/spatial-transformer-network.

STN链接: https://github.com/Henuzhaoyli/STN_in_pytorch.

1 摘要

很多方法都用tripletloss研究行人重识别,但是行人的姿态变化和视角限制了算法。为了解决这个问题,作者引入了STN(spatial transformer network)来对齐行人。
贡献:引入STN模块在行人重识别领域,(但并不是第一个)减轻姿态变换带来的影响。
对VGG、ResNet和DenseNet三种网络体系结构进行了评估,经验性地说明了STN模块的通用性。在实验结果的基础上,我们提出了一个由STN模块、DenseNet骨干网和TriHard损耗组成的鲁棒、高性能的REID模型。
通过公式推导,证明了我们的模型是可微的,从而实现了端到端的Reid系统。

2 理论研究

Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
与ResNet相比,DenseNet充分利用了中间层信息丰富的特点。此外,DenseNet在参数效率方面具有优势,并改善了网络中的信息流和梯度。DenseNet不仅具有优于ResNet的性能,而且体现了STN模块在更深层次架构上的通用性

2.1 STN (spatial transform net)

Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录

Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录
第二幅图片需要修改 维度,因为我是用cifar-10,彩色图像,大小尺寸也不一样,第一幅用的MNIST数据集。

STN知识参考链接: https://www.cnblogs.com/liaohuiqiang/p/9226335.html.

3 实验结果

Deeply-Learned Spatial Alignment for Person Re-Identification 论文阅读记录