3D Face Reconstruction from a Single Image

Michael Hobbel (Utrecht University)

Abstract

This report presents the conclusions of the experimentation project performed by Utrecht University master student Michael Hobbel under supervision of dr. Robby T. Tan. The goal of the project was to determine the possible drawbacks of the 3D face reconstruction method described in Face Reconstruction From a Single Image Using a Single Reference Face Shape (Kemelmacher-Shlizerman & Basri, 2009). The method exploits the general similarity in faces by using a single reference model as a base and molding it towards a specific face by using shape-from-shading on the input image. It allows (possibly multiple) unknown lighting conditions, as it assumes the face to be Lambertian and it models the reflection function using the first order spherical harmonics, this makes the method both fast and robust. Then finally, once the lighting is known and the face's surface has been reconstructed using shape-from-shading, the albedo is calculated. I have made an implementation of this method in c++, to generate results that could confirm or refute hypothetical drawbacks and then conducted a survey to prove these hypotheses quantitatively.


Report

3D Face Reconstruction from a Single Image [PDF]
Michael Hobbel


System Pipeline


Results:

For each input image: A projection of the constructed 3D model with a constant albedo, a projection of the constructed 3D model with the estimated albedo and a projection of the constructed 3D model with the input image as overlay (not lighted in this case)

References