src/pixel3dmm/preprocessing/MICA/datasets/README.md

MICA - Dataset

The MICA dataset consists of eight smaller datasets for about 2315 subjects, built by unifying existing small- and medium-scale datasets under a common FLAME topology. It consists of shape geometry only, therefore, to obtain images for each subject please refer to the primary dataset.

This dataset contains registration meshes together with corresponding fitted FLAME parameters. Actors are split between individual folders with a unique identifier based on the original dataset. The folder name of the parameters and mesh is the same as the ones in the analogous dataset with images.

To obtain the dataset please follow each link separately and request the given subset.

In the case of any questions feel free to email us.

FLAME Renderings	Dataset	Subjects	Source
	Stirling	133	pics.stir.ac.uk
	D3DFACS	10	cs.bath.ac.uk
	Florence 2D/3D	53	micc.unifi.it
	LYHM	1211	cs.york.ac.uk
	FaceWarehouse	150	kunzhou.net
	*FRGC	531	cvrl.nd.edu
	*BP4D+	127	cs.binghamton.edu

Each subset zip file has the following structure:

root\
    FLAME_parameters\
        actor_id\
            *.npz
    registrations\
        actor_id\
            *.obj

To retrieve FLAME2020 parameters you can simply do:

import numpy as np
import torch

params = np.load('path.npz', allow_pickle=True)
pose = torch.tensor(params['pose']).float()
betas = torch.tensor(params['betas']).float()

flame = {
    'shape_params': betas[:300],
    'expression_params': betas[300:],
    'pose_params': torch.cat([pose[:3], pose[6:9]]),
}

MICA Training Dataset Preparation

To prepare the MICA training dataset you can follow the scripts from the creation folder. Additionally, the complete list of images used for the training can be found in image_paths folder. It contains the name of the FLAME parameters file .npz and a list of all images used for the training with their original name stored as dictionary.

Licence

This dataset is for academic, non-commercial usage only. Moreover, it is an extension of already existing datasets, therefore, the license is shared and applies equivalently to both, the original and the corresponding derived one. Please read the original license of each original dataset for more information, especially in the context of data privacy.

Additional Information

Please note that some subsets do not have an identical amount of subjects due to registration errors.

BP4D+ Subset

The original dataset contains 140 subjects from where we selected scans in neutral pose and successfully registered 127 FLAME meshes for them.

FRGC Subset

1. A group desiring to obtain FLAME results on the FRGC data must license the FRGC 2.0 data set from CVRL using the licensing procedure at the website: http://cvrl.nd.edu .
2. Once the license form is completed and CVRL has approved it, an email will be sent from Globus indicating that you have been authorized to retrieve FRGC 2.0.
3. That authorization will serve as authorization to receive any derivative work, therefore, please send a confirmation to mica [AT] tue.mpg.de address to receive the MICA dataset.

Citation

If you use this dataset 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}
}