1. Introduction
- Attribute: A meaningful feature inherent in an image such as hair color, gender or age.
- Attribute value: A particular value of an attribute, e.g., black/blond/brown for hair color or male/female for gender.
- Domain: A set of images sharing the same attribute value.
- Multi-domain image-to-image translation: We change images according to attributes from multiple domains.
- Training multiple domains from different datasets is possible, such as jointly training CelebA and RaFD images to change a CelebA image's facial expression using features learned by training on RaFD.
- hyeonjon modeldeulyi munjejeom:
-
$k$gaeyi Domaini isseul ddae ${}{k}P{2}$gaeyi modeli pilyohabnida.
- jeonce Domaine daehan gongtongdoen Featuresga issdago hadeorago 2gaeyi Domaineurobuteobagge hagseubi bulgahabnida.
- modeleun Domainyi jeongboreul One-hot encoded labelroseo badadeulibnida. hagseub jungeneun Target domain labeleul mujagwiro jeonghago geu Domaineuro imijireul beonyeoghadorog modeli hagseubdoebnida.
- Adding a mask vector to the domain label:
- moreuneun Labeleul musihago teugjeong deiteosese yihae jueojineun Labeleman jibjunghadorog habnida.
2. Related Work
3. Star Generative Adversarial Networks
3.1) MultiDomain Image-to-Image Translation
-
$x$: Source domain image.
-
$y$: Target domain image.
-
$c$: Randomly generated target domain label.
$$G(x, c) \rightarrow y$$
-
$D_{\text{src}}(x)$: Probability distribution over sources.