Face synthesis from a frontal pose image using partial least squares and b-splines

This paper investigates the effect of b-splines as a representation for dense 3D facial shape estimation under the partial least squares regression scheme. To this end, a frontal grayscale image is first used for predicting a texture map expressed in cylindrical coordinates. The texture map is then used for estimating a sparse grid of control points from which a b-spline surface is constructed. The prediction steps are carried out through a regression matrix built from the partial least squares technique. Experimental results demonstrate that using b-splines considerably reduces the information contained in a dense 3D cylindrical map, while efficiently preserving the accuracy provided by cylindrical dense maps for the purposes of statistical 3D shape estimation.