Date: June 2019.
Source: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA.
Research Summary: Earlier methods could reconstruct 3D face shape from a single photo, but the facial texture they produced was blurry and lacked realistic skin detail. GANFIT solves this with a generative adversarial network, or GAN, trained directly on high-resolution facial texture maps. The method fits this GAN, together with a 3D morphable model, to a single input photo, guided by a face recognition network to preserve the person’s identity. The result is a 3D reconstruction with realistic texture detail and an accurate likeness, outperforming earlier methods on standard 3D face benchmarks.
3dMD’s Role: GANFIT’s 3D morphable model is built from LSFM, which itself comes from 3dMD’s MeIn3D dataset. GANFIT’s GAN is trained on roughly 10,000 high-resolution facial texture maps, also drawn from MeIn3D. GANFIT’s expression model is trained separately, on the 4DFAB facial expression database.

Article: GANFIT: Generative Adversarial Network Fitting for High Fidelity 3D Face Reconstruction.
Authors: B Gecer, S Ploumpis, I Kotsia, S Zafeiriou. Imperial College London.

Related Articles:

  1. Handy: Towards a high fidelity 3D hand shape and appearance model. R Potamias, S Ploumpis, S Moschoglou, V Triantafyllou and S Zafeiriou.
  2. A generative model with domain adversarial training and independent conditional discriminators to predict the development of craniosynostosis. C Elkhill, IA Cruz-Guerrero, MG Linguraru, B French, AR Porras.
  3. 3D human tongue reconstruction from single “in-the-wild” images. S Ploumpis, S Moschoglou, V Triantafyllou, S Zafeiriou.
  4. ImHead: A Large-scale Implicit Morphable Model for Localized Head Modeling. RA Potamias, S Galanakis, J Deng, A Papaioannou, S Zafeiriou.