Public repository for the CVPR 2020 paper AvatarMe, with high resolution results, data and more.
Alexandros Lattas 1,2,
Stylianos Moschoglou 1,2,
Baris Gecer 1,2,
Stylianos Ploumpis 1,2,
Vasileios Triantafyllou 2,
Abhijeet Ghosh 1,
Stefanos Zafeiriou 1,2
1 Imperial College London
2 FaceSoft
AvatarMe is the first method that is able to reconstruct photorealistic 3D faces from a single ‘in-the-wild” image with an increasing level of detail. To achieve this, we capture a large dataset of facial shape and reflectance and build on a state-of-the 3D texture and shape reconstruction method and successively refine its results in order to generate the high-resolution diffuse and specular components that are required for realistic rendering.
A 3DMM is fitted to an ''in-the-wild'' input image and a completed UV texture is synthesized, while optimizing for the identity match between the rendering and the input. The texture is up-sampled 8 times, to synthesize plausible high-frequency details. We then use an image translation network to de-light the texture and obtain the diffuse albedo with high-frequency details. Then, separate networks are used to infer the specular albedo, diffuse normals and specular normals from the diffuse albedo and the 3DMM shape normals. Moreover, the networks are trained on 512x512 patches and inferences are ran on 1536x1536 patches with a sliding window. Finally, we can transfer the facial shape and the consistently inferred reflectance to a head model. Both face and head can be rendered realistically in any environment.
Patch process for Diffuse and Specular Abedo
Patch process for Diffuse and Specular Normals
For high quality rendered results see the following youtube video.
RealfaceDB can now be obtained by accredited researchers and used for training/testing methods similar to AvatarMe.
The dataset contains patches of facial reflectance as described in the paper,
namely the diffuse albedo, diffuse normals, specular albedo, specular normals,
as well as the shape in UV space. For the shape,
reconstructed meshes have been registered to a common topology
and the XYZ values of the points have been mapped to the RGB in UV coordinates and interpolated to complete the UV map.
From the complete UV maps of 6144x4096 pixels, patches of 512x512 pixels have been sampled.
The dataset contains 7500 such patches (1500 of each datatype) that are anonymized, randomized
and sampled so that they do not contain identifiable features.
To obtain access to the dataset, you need to complete and sign a licence agreement, which should be completed by a full-time academic staff member (not a student). To obtain the licence agreement and the dataset please send an email to Alexandros Lattas (a.lattas@imperial.ac.uk) and Stylianos Moschoglou (s.moschoglou@imperial.ac.uk). Please contact us through your academic email and include your name and position. We will verify your request and contact you regarding how to download the dataset. Note that the agreement requires that:
- The data must be used for non-commercial research and education purposes only.
- You agree not to copy, sell, trade, or exploit the model for any commercial purposes.
- You must destroy the data after 2 years since the first download.
- If you will be publishing any work using this dataset, please cite the following paper.
If you find this work useful, please use the following to cite our paper:
@InProceedings{Lattas_2020_CVPR,
author = {Lattas, Alexandros and Moschoglou, Stylianos and Gecer, Baris and Ploumpis, Stylianos and Triantafyllou, Vasileios and Ghosh, Abhijeet and Zafeiriou, Stefanos},
title = {AvatarMe: Realistically Renderable 3D Facial Reconstruction "In-the-Wild"},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}