An Efficient Markerless Method for Resynthesizing Facial Animation on an Anatomy-Based Model

This paper describes a markerless method for automatically estimating animation parameters of an anatomy-based face model from a video sequence. The method is based upon an efficient tracking procedure which employs a linear subspace model of facial appearance. During the training phase of the tracking algorithm, a set of templates is computed from the subspace basis to alleviate the online computation. The efficient tracking algorithm establishes temporal correspondence of the face region by simultaneously determining both motion and appearance parameters. Using a set of key expressions of the actor together with corresponding animation parameters of the face model, we learn a linear relationship between facial animation parameters and appearance parameters. It allows us to efficiently estimate the animation parameters from appearance parameters provided by the tracker. We present experimental results performed on synthetic and live video sequences to demonstrate the effectiveness and efficiency of our method.