Data-Driven Facial Feature Morphing for 3D Face Synthesis

Abstract

This paper presents a data-driven method for synthesizing new realistic 3D human faces by morphing a set of facial features. The method takes as examples 3D scanned human face models in order to exploit the shape variations presented in the real faces of individuals. We automatically compute a vertex-to-vertex correspondence between the unregistered face scans by deforming a generic mesh to fit the specific person’s face geometry in a global-tolocal fashion. Exploiting the statistics of the generated datasets of feature shapes we transform them into vector space representations by applying a principal component analysis (PCA). Our feature shape morphing model is formed as a linear combination of the main modes of feature shape variations. We introduce a shape smoothing and blending method for generating a seamlessly deformed mesh around the feature borders. We demonstrate our method by combining 3D morphing of four features to generate a wide range of different face shapes.