/ Computer Vision

Image Translation with GAN (1)

Comments

Problem statement of Image Translation

inline

Learn $G: (S \rightarrow T)$

$G$ that convert an image of source domain $S$ to an image of target domain $T$
Domain Adaptation/Transfer of an image.

Paired and Unpaired dataset

inline

Image Translation: $S$ and $T$ are pair-wise labeled in dataset

inline

Image Translation: $S$ and $T$ are not pair-wised in dataset

inline

fit
inline

Before, Style Transfer (NeuralArt) was prominent

But it largely depends on textual information of an target style

inline fit
inline fit

How to learn more general Image Translation?

Generative Adversarial Network

fit right
inline
inline

\begin{equation}
\min_G \max_D V(D,G) = \mathbf{E}_{x\sim p_{\text{data}}(x)} [\log D(x)] + \mathbf{E}_{z\sim p_{z}}(z)[\log (1 - D(G(z)))].
\end{equation}

also honorable mention : Deep Convolutional GAN (DCGAN)

Two major problems of Image Translation

  1. Convert to which domain?
    • learn which "$G: (S \rightarrow T)$"?
  2. How to learn the dataset?
    • how to properly form dataset?
    • pair-wise Supervised? or Unsupervised?

Today, presenting SOTA of Image Translation papers of

  • pix2pix: Image-to-Image Translation with Conditional Adversarial Networks (CVPR2017)
  • Domain Transfer Network: Unsupervised Cross-Domain Image Generation (ICLR2017)
  • CycleGAN & DiscoGAN: CycleGAN: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (ICCV2017) & DiscoGAN: Learning to Discover Cross-Domain Relations with Generative Adversarial Networks (ICML2017)
  • BicycleGAN: Toward Multimodal Image-to-Image Translation (NIPS2017)
Junho Cho

Junho Cho

Integrated Ph.D course and Interested in Computer Vision, Deep Learning. For more information, tmmse.xyz/junhocho/

Read More