Example-based Learning for Depth Estimation ofFacial Landmarks and Its Application in FaceModeling

When reconstructing a 3D face from a singleimage, the unknown depth of landmarks degrades themodel’s accuracy considerably. A promising solution is topredict the depth of landmarks by learning from 3D examplesof scan before modeling. This paper proposed to usea sparse linear model to estimate the depth of landmarksfrom their prior distributions in a 3D face database. Theestimated 3D landmarks were applied to the deformationprocess to ensure a more precise facial shape for a given image.Tests on synthesized images show that the estimatedfeatures’ depth is closely to the known ground truth andthat the modeling accuracy of various deforming methodsis greatly enhanced with the estimated 3D features.