Arxiv 2021: Accurate 3D Facial Geometry Prediction by Multi-Task, Multi-Modal, and Multi-Representation Landmark Refinement Network
Cho-Ying Wu, Qiangeng Xu, Ulrich Neumann, CGIT Lab at University of Souther California
[paper] [project page]
Advantages:
+ State-of-the-art on all 3D facial alignment, face orientation estimation, and 3D face modeling.
+ Exploitation of Multi-Modal and Multi-Representation for information aggregation.
+ Fast and easy to use: 3000fps for 3D facial landmarks on a single GPU.
+ The first to study the face reconstrcutability from sparse landmarks.
More results:
Facial alignemnt on AFLW2000-3D (NME of facial landmarks):
Face orientation estimation on AFLW2000-3D (MAE of Euler angles):
General results:
Updates: The codes are now released https://github.com/choyingw/SynergyNet. This repo will be archived due to its older version.