Conference Paper/Proceeding/Abstract 231 views
One-Shot Decoupled Face Reenactment with Vision Transformer
Pattern Recognition and Artificial Intelligence, Volume: Lecture Notes in Computer Science (LNCS, volume 13364), Pages: 246 - 257
Full text not available from this repository: check for access using links below.
Recent face reenactment paradigm involves estimating an optical flow to warp the source image or its feature maps such that pixel values can be sampled to generate the reenacted image. We propose a one-shot framework in which the reenactment of the overall face and individual landmarks are decoupled...
|Published in:||Pattern Recognition and Artificial Intelligence|
Springer International Publishing
Check full text
No Tags, Be the first to tag this record!
Recent face reenactment paradigm involves estimating an optical flow to warp the source image or its feature maps such that pixel values can be sampled to generate the reenacted image. We propose a one-shot framework in which the reenactment of the overall face and individual landmarks are decoupled. We show that a shallow Vision Transformer can effectively estimate optical flow without much parameters and training data. When reenacting different identities, our method remedies previous conditional generator based method’s inability to preserve identities in reenacted images. To address the identity preserving problem in face reenactment, we model landmark coordinate transformation as a style transfer problem, yielding further improvement on preserving the source image’s identity in the reenacted image. Our method achieves the lower head pose error on the CelebV dataset while obtaining competitive results in identity preserving and expression accuracy.
ICPRAI 2022. Lecture Notes in Computer Science, vol 13364..
Faculty of Science and Engineering