Image to Image Translation for Domain Adaptation(CVPR2018)

# Background

Doing the same Domain Adaptation job as CYCADA, translate images between GTA5 and Cityscapes.

This is published after CYCADA.

# Method

The main difference is that there's a Embedding space in the middle.

![title](https://leanote.com/api/file/getImage?fileId=5d89ebebab6441211d001a30)

The loss consis of 6 parts:

* Normal Classfication (P1)

![title](https://leanote.com/api/file/getImage?fileId=5d89f375ab6441211d001a9c)

* Reconstruct part (P2)

![title](https://leanote.com/api/file/getImage?fileId=5d89ed7dab64411f16001b02)

* Discriminator z part (P3)

![title](https://leanote.com/api/file/getImage?fileId=5d89ede3ab6441211d001a42)

* Discriminator x,y part (P4)

![title](https://leanote.com/api/file/getImage?fileId=5d89eedbab64411f16001b0f)

* Cycle loss (P5)

![title](https://leanote.com/api/file/getImage?fileId=5d89ef91ab64411f16001b19)

* Complex Classfication (P6)

![title](https://leanote.com/api/file/getImage?fileId=5d89f361ab6441211d001a9a)

In total, there's 3 discriminator $D_z,D_x,D_y$ and 2 encoder $f_x,f_y$ and 2 decoder $g_x,g_y$ and a classifier $h$.

# Performance

State of art performance when this paper was published.

![title](https://leanote.com/api/file/getImage?fileId=5d89f617ab64411f16001b54)

In job above the encoder is LeNet.

![title](https://leanote.com/api/file/getImage?fileId=5d89e9abab64411f16001ad2)

Someting interesting: "Switching to a DenseNet encoder beats the previous method **even without domain adaptation"**

CYCADA use CycleGAN who only used Resnet or Unet. And CYCADA has a sematic consistent loss.

![title](https://leanote.com/api/file/getImage?fileId=5d89f3e2ab64411f16001b3f)

# Inspiration

This paper tells us how to do Domain Adaptation across multi source. Combined with semi-Supervised learning may have better perfomance.

wuvin