/MICA

MICA - Towards Metrical Reconstruction of Human Faces [ECCV2022]

Primary LanguagePythonOtherNOASSERTION

MICA - Towards Metrical Reconstruction of Human Faces

Wojciech Zielonka, Timo Bolkart, Justus Thies

Max Planck Institute for Intelligent Systems, Tübingen, Germany

Video  Paper  Project Website  Face Tracker  Dataset  Supplemental  Email

Official Repository for ECCV 2022 paper Towards Metrical Reconstruction of Human Faces

⚠ The face tracker is now available under Metrical Photometric Tracker  ⚠

Installation

After cloning the repository please install the environment by using attached conda environment.yml file with the command conda env create -f environment.yml. Additionally, the FLAME2020 model is needed. To obtain it please create an account at the website download the model and place it in the /data/pretrained/FLAME2020/ folder.

You can also simply run the install.sh script:

git clone https://github.com/Zielon/MICA.git
cd MICA
./install.sh

you will be asked to provide {flame_user} and {flame_password} for your FLAME account in order to access the file server.

Pre-trained Models

If you decide to not use the installation script, the pretrained model can be found under the MPI-IS storage server. After downloading, please place it in the /data/pretrained/mica.tar location. Additionally, you will need to provide models for inisghtface:

  1. antelopev2
  2. buffalo_l

then you need to unzip them and place in ~/.insightface/models/. The install.sh script does it for you.

How To Use

To use MICA you can simply run the demo.py file. It will process all the images from demo/input/ folder and create the output destination for each subject with .ply mesh, rendered image, and .npy FLAME parameters.

Dataset and Training

The MICA dataset consists of eight smaller datasets for about 2300 subjects under a common FLAME topology. Read more information about how to obtain and use it under the link. To train MICA the images from all eight datasets are needed. The repository contains scripts how to generate the Arcface input images as well as the complete list of all the images used for the training. More information can be found here.

When you train from scratch for Arcface model initialization please download Glint360K and specify the path to it in the config as cfg.model.arcface_pretrained_model.

Testing

The testing was done using two datasets, Stirling and NoW. In the model folder you can find the corresponding scripts to run testing routine, which generates the meshes. To calculate the NoW challenge error you can use the following repository.

Citation

If you use this project in your research please cite MICA:

@proceedings{MICA:ECCV2022,
  author = {Zielonka, Wojciech and Bolkart, Timo and Thies, Justus},
  title = {Towards Metrical Reconstruction of Human Faces},
  journal = {European Conference on Computer Vision},
  year = {2022}
}