{%hackmd @themes/dracula %}

==Problem==
照片有些細節無法被生成出來，ex. 雨滴的掉落

## method

==Domain bridging==

一種資料擴增的方式

$X_W, Y_W$: related with weather domain
$X_O, Y_O$: unrelated with weather domain

如果兩個資料集的差異性最小化->穩定的I2I transformer

Reduce the KL divergence between $P_{X'_O}$ and $P_{Y'_O}$, with respect to $KL(P_{X_O}(x), P_{Y_O}(y))$

**How to choose?**

$max|Z^x_W\cap X_W|$
$max|Z^y_W\cap Y_W|$
$max|Z^x_O\cap Z^y_O|$

==Unsupervised Domain Adaptation (UDA)==

==Online Multimodal Style-sampling (OMS)==

From MUNIT, it can generate multiple target styles(rain appearances)

==Weighted Pseudo-Labels (WPL)==