GANimation: Anatomically-aware Facial Animation from a Single Image
Continuous changing of AUs using unsupervised GAN.
Formally, it learns a model
DIAT, CycleGAN, IcGAN, StarGAN
Losses (four terms)
- traditional GAN loss: try to make the generated image as real as possible
- Attention loss: keep the mask smooth (TV) and force the mask to use the generated color map (L2)
- Conditional Expression Loss: training a separate AU regressor
- Identity Loss: nothing but a cycle loss
Dataset is EmotioNet dataset, result is amazing. The model could generate continuous AU changes, which is what I currently working on, SAD…
To tell the truth, this paper does not introduce really new stuff, WGAN-GP, TV loss, extra classifier/regressor, identity (cycle) loss, these are all techniques that have been used before, the good part is they combine them into a one single model and get outstanding performance! Excellent job!