Robust 3D face reconstruction: A "movement-structure" dual approach

An original approach for three-dimensional reconstruction and animation of human face with a monocular and un-calibrated acquisition system is present ed This method is based on the classical analysis-synthesis scheme. A three -dimensional reconstruction of shape and movement is performed with a combi nation of two algorithms using the well-known "Z from motion" principle. Th erefore a pre-processing step is needed to first detect the human face on v ideoconference sequence. then, a set of points is tracked with a new hybrid and adaptive block-matching algorithm using the Delaunay tessellation. A l ocal algebraic projective invariant family performs the 3D movement classif ication to satisfy the 3D synthesis assumptions. Experimental results on sy nthetic and noisy data set validate the entire approach.