Training a Neural Net in Pytorch to generate 3d facial pose estimation from only facial landmarks.
This is one of three repositories which together form a larger project, these are the three repositories:
- facial-pose-estimation-pytorch-v2 - You are here.
- facial-pose-estimation-libtorch-dll
- facial-pose-estimation-unreal
This blog post describes the whole project: AR Facial Pose Estimation
- Training code is in Jupyter Notebook
- Training data is artificially generated on the fly using OpenCV
- Output is 51 blend-shape float values which can be applied to a face mesh
- This only outputs blend-shape values, head transform is handled by a separate part of the pipeline
Usage instructions found here: user manual page.