You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
현존 모델들의 문제점:
$k$개의 Domain이 있을 때 ${}{k}P{2}$개의 모델이 필요합니다.
전체 Domain에 대한 공통된 Features가 있다고 하더라고 2개의 Domain으로부터밖에 학습이 불가합니다.
모델은 Domain의 정보를 One-hot encoded label로서 받아들입니다. 학습 중에는 Target domain label을 무작위로 정하고 그 Domain으로 이미지를 번역하도록 모델이 학습됩니다.
Adding a mask vector to the domain label:
모르는 Label을 무시하고 특정 데이터셋에 의해 주어지는 Label에만 집중하도록 합니다.
2. Related Work
생략합니다.
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.
About
PyTorch implementation of 'StarGAN' (Yunjey Choi et al., 2018)