The GeoGen dataset is a comprehensive collection of over 70,000 synthetic face images designed for advanced 3D geometry reconstruction research.
The dataset is essential for training deep learning models that are geared towards high-fidelity 3D facial geometry reconstruction.
This dataset was introduced in our paper titled GeoGen: Geometry-Aware Generative Modeling via Signed Distance Functions.
The dataset contains:
- The dataset comprises 70,000 images distributed across 10,000 identities, with each identity represented by 7 unique images. These images are captured from various camera angles spanning a full 360-degree view to enhance the diversity and comprehensiveness of the dataset. The dataset includes the camera extrinsics and intrinsics in json format.
The GeoGen dataset can be used for non-commercial research, and is licensed under the license found in LICENSE.
For convenience the images dataset is split into 7 parts and the last part contains the metadata with the camera parameters which can be downloaded here:
7 images per identity
Camera parameters:
The GeoGen dataset contains cropped color images in the following layout. For the camera parameters the average focal lenght is 50 and the sensor width is 36.
subj_id_n
├── 0.png # First rendered image of subject subj_id_n
├── 1.png # Second rendered image of subject subj_id_n
...
├── k.png # k+1 rendered image of subject subj_id_n
metadata_id_n_jsom # The extrinsics and intrinsics are in their respective metadata.json files in the following layout.
├── name of the subject # Corresponding name image from subj_id_n
├── cameras # Camera parameters of image from subj_id_n
Some of our rendered faces may be close in appearance to the faces of real people. Any such similarity is naturally unintentional, as it would be in a dataset of real images, where people may appear similar to others unknown to them.
If you use the GeoGen dataset in your work, please cite the following paper:
@inproceedings{esposito2024geogen,
author = {Esposito, Salvatore and Xu, Qingshan and Kania, Kacper and Hewitt, Charlie and Mariotti, Octave and Petikam, Lohit and Valentin, Julien and Onken, Arno and Mac Aodha, Oisin},
title = {GeoGen: Geometry-Aware Generative Modeling via Signed Distance Functions},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2024},
pages = {7479-7488}
}