2017 CVPR

官方代码：https://www.uow.edu.au/~jz960/

读完，感觉以下两点思想感觉和2018年的DICD有点相似之处：

1. the variance of target domain is maximized,
2. the discriminative information of source domain is preserved

但是DICD考虑的是作域变换之后数据之间的可分性，同类别距离小，不同类别距离大。

1 introduction

In this paper, we propose a unified framework that reduces the distributional and geometrical divergence between domains simultaneously by exploiting both the shared and domain specific features.

Specifically, we learn two coupled projections to map the source and target data into respective subspaces.

After the projections,

1. the variance of target domain data is maximized to preserve the target domain data properties,
2. the discriminative infor- mation of source data is preserved to effectively transfer the class information,
3. both the marginal and conditional distribution divergences between source and target domains are minimized to reduce the domain shift statistically, 
4. the divergence of two projections is constrained to be small to reduce domain shift geometrically.

2 related work

3 Joint Geometrical and Statistical Alignment

The JGSA is formulated by finding two coupled projections (A for source domain, and B for target domain) to obtain new representations of respective domains, such that

1. the variance of target domain is maximized,
2. the discriminative information of source domain is preserved,
3. the divergence of source and target distributions is small,
4. the divergence between source and target subspaces is small.

将下列（1）（3）（4）（9）（14）公式合并到一起。

3.2.1 Target Variance Maximization

3.2.2 Source Discriminative Information Preservation

3.2.3 Distribution Divergence Minimization

就是JDA的公式

3.2.4 Subspace Divergence Minimization

3.2.5 Overall Objective Function

For JGSA, we fix λ = 1, μ = 1 in all the experiments, such that the distribution shift, subspace shift, and target variance are treated as equally important. 

3.3. Optimization

最后求解公式（19），得到W。

4. Experiments

实验结果