Learning Based Prediction and 3D-Visualisation of Children's Facial Growth

Based on a novel, example-based approach for studying how children’s faces change as they grow, we present an automated algorithm for the prediction of children’s facial growth. A single image of an arbitrary face at its present appearance is sufficient to estimate an age-progressed 3D face model of the individual. By extracting and applying growth vectors from a previously acquired database, the face can be transformed into its individual older or younger appearance. The database is a combination of an existing set of 3D laser scans of 200 adult human faces and two newly acquired 3D databases of baby and teenager faces. Those are generated from 3D laser scans and stereo images, respectively. By incorporating the data into a morphable 3D face model, a homogeneous face space is constructed, which contains 523 3D face models of adults, teenagers and babies in a vector representation that involves dense point-to-point correspondence between individual faces. For the baby face models, additional anthropometric measurements of the facial soft tissue are performed by a novel, automated algorithm. The results provide a longitudinal study of baby faces, which is intended to assist in detecting anomalies of the facial growth at early stages of development. The additional shape and texture vectors for children and a subsequent Principal Component Analysis improve the reconstruction results of the morphable 3D face model when matching it to infant faces. Moreover, a subdivision of the database into age groups makes it possible to describe the individual growth of a face continuously, starting from a minimum age of 3 months to a maximum age of 35 years. In the vector space representation, growth is modelled by a piecewise linear function that approximates a desired target age for an individual face by learning the growth from the given sample set of 523 faces.