Leveraging SSL 2 Generate High Quality 3D Face Avatar from Portrait Image

High-fidelity 3D face reconstruction from monocular images aims to obtain a 3D representation of the subject from a single or multiple input image. Recently, self-supervised deep-based methods have demonstrated impressive performance in 3D face reconstruction. These methods are efficient and produce plausible face reconstruction. However, for AAA production (games and movies), they do not yet meet the production-level requirements. For instance, the estimated geometry does not fully recover the likeness of the subject and the estimated texture maps used for rendering typically have low resolution. However, at least 4K texture maps are required in AAA productions. In this work, we aim to push the quality of the reconstruction provided by self-supervised methods to reach production-level needs.

Faculty Supervisor:

Steve Engels

Student:

Partner:

Ubisoft Toronto

Discipline:

Computer science

Sector:

Information and cultural industries

University:

University of Toronto

Program: