Abstract: The thin-shell deformation algorithm calculates the displacements of all the vertices based on the displacements of control points.The reconstruction results may be different by using different sets of control points.An algorithm to extract optimal control points based on thin-shell method is proposed in order to reconstruct the facial motion data more accurately.First,this problem is formalized as minimizing the reconstruction.Then an adaptive forward-backward greedy strategy is proposed to learn the optimal control points which are called key points in our work.The boundary constraint,symmetry constraint and multi-resolution property are introduced and with these considerations,result key points are very robust.The experiments and comparisons demonstrate the proposed optimal key points could be used for effectively reconstructing facial animation.