Abstract
Near all people’s faces are different from each other, except perhaps for identical twins. But what defines this difference? To a large part, our physical appearance is genetically determined, which suggests that there must be specific patterns of variation to our appearance. This is also reflected in our ability to compare faces in manners such as: “she has the eyes of her mother”. Despite the apparent patterns of differences and similarities in faces, quantification of such traits has not been extensively performed. The recent development of 3D capturing devices, together with 3D models for dense surface registration now make it possible to systematically compare large collections of faces and discover latent features hidden behind our appearance. That has been the topic of research presented in this thesis. We investigate machine learning approaches for 3D face mensuration, with the aim to reveal genetically determined facial features. This involves establishing a measurement system for the consistent extraction of information from large volumes of 3D photos, and a measurement analysis based on this, to learn phenotypic variants. Consistent extraction of 3D shape information can be performed using a 3D shape model that, by matching it to new images, conforms the photographed facial surface to a fixed set of measured points. The approach depends on an initial set of manually marked reference points. Locating these points, however, becomes unpractical for large amounts of images. In order to deal with this, we therefore develop a method that can infer the positions of those reference points in new images automatically. The measurement analysis is based on defining a set of coherent parts whose shapes exhibit a large genetic influence. Such parts can be considered as latent factors in the face shape space. Traditional methods for learning latent factors from data result in holistic features, describing shape variation across the whole face. A more recent technique is non-negative matrix factorisation, which has shown to result in sparser, and also more local, features. Parts learnt using this approach define coherent regions of shape variation. We study the shape variation in each part, as well as combinations of parts, using an estimate for the amount of genetic influence to pick candidate features that will be subjected to gene association in a future study.