Generating 3D Virtual Populations from Pictures of a Few Individuals

This paper describes a method for cloning faces from two orthogonal pictures and for generating populations from a small number of these clones. An e cient method for reconstructing 3D heads suitable for animation from pictures starts with the extraction of feature points from the orthogonal picture sets. Data from several such heads serve to statistically infer the parameters of the multivariate probability distribution characterizing a hypothetical population of heads. A previously constructed, animation-ready generic model is transformed to each individualized head based on the features either extracted from the orthogonal pictures or determined by a sample point from the multivariate distribution. Using projections of the 3D heads, 2D texture images are obtained for individuals reconstructed from pictures, which are then tted to the clone, a fully automated procedure resulting in 360 texture mapping. For heads generated through population sampling, a texture morphing algorithm generates new texture mappings.