MASTER Assignment
creating 3d faces from 2d images using gan s
Type : Master M-CS
Period: Nov, 2019- Jun, 2020
Student : Reesink, T.B.J. (Thomas, Student M-CS)
Date Final project: June, 29, 2020
Supervisors:
Abstract:
This research looks into the viability of generating realistic 3D data based on a single 2D image by means of Generative Adversarial Networks. Many existing methods for generating 3D faces from 2D information require models, making them lose detail as they tend to smooth identifying traits. This research aims to make a system which generates raw 3D data and does not require a predefined model. This is achieved by utilizing Generative Adversarial Networks (GANs) which can generate convincing samples based on a given dataset. By conditioning the GAN it is possible to base the generated 3D data on a given 2D image. In order to objectively measure the quality of 3D data generated by its models, this research trains its models using 3D facial data and uses 3D facial recognition to verify its results. Using 3D facial recognition on the generated samples allows for comparison between methods as well as new insights. By transforming 3D to a 2D matrix it was possible to train a conditional Wasserstein GAN to produce 3D data which could be correctly identified in 63.3% of the cases.