Introduction

This repository refer to 3DDFA and DECA.

3DDFA use 40 shape parameter and 10 expression parameter. In order to improve the reconstruction accuracy, this repository try to use 199 shape parameter and 29 expression parameter. but the result don't seem to improve much. (dataset:300W_LP, net: mobilenet_v1)

DECA has better result in close eye and the outline of the mouth. this repository just extract the 3d facial landmark detection code.

Requirements

• Python 3.7 (numpy, skimage, scipy, opencv)
• PyTorch >= 1.6 (pytorch3d)
  You can run

  pip install -r requirements.txt

  Then follow the instruction to install pytorch3d.

Usage

1. Download config(3DDFA) and model

• config file Google Cloud
• related model Google Cloud

2. Run demos

3DDFA

python detect_landmark_bfm.py video/demo.mp4

DECA

python detect_landmark_flame.py video/demo.mp4

facial landmark face-alignment

License

The code is available for non-commercial scientific research purposes.

Notes

We use the FLAME.py from FLAME_PyTorch