General Human Traits Oriented Generic Elastic Model for 3D Face Reconstruction

We propose a Simplified Generic Elastic Model (S-GEM) which intends to construct a 3D face from a given 2D face image by making use of a set of general human traits viz., Gender, Ethnicity and Age (GEA). Different from the original GEM model which employs and deforms the mean depth value of 3D sample faces according to a specific 2D input face image, we hypothesise that the variations inherent on the depth information for individuals are significantly mitigated by narrowing down the target information via a selection of specific GEA traits. This is achieved by representing the unknown 3D facial feature points of a 2D input as a Gaussian Mixture Model (GMM) of that of the samples of its own GEA type. It is then further incorporated into a Bayesian framework whereby the 3D face reconstruction is posed as estimating the PCA coefficients of a statistical 3D face model, given the observation of 2D feature points, however, with their respective depth as hidden variables. By making the reasonable assumption that the support area of each component of GMM is small enough, the proposed method is reduced to choose the depth values of the features points of a sample face that is nearest to the 2D input face. Thus the 3D reconstruction is obtained with depth-value augmented feature points rather than the 2D ones in normal PCA statistical model based reconstruction. The proposed method has been tested with the USF 3D face database as well as the FRGC dataset. The experimental results show that the proposed S-GEM has achieved improved reconstruction accuracy, consistency, and the robustness over the conventional PCA based and the GEM (mean-face feature points) reconstruction, and also yields enhanced visual improvements on certain facial features.