Realistic face animation from sparse stereo meshes

Being able to produce realistic facial animation is crucial for many speech applications in language learning technologies. For reaching realism, it is necessary to acquire and to animate dense 3D models of the face. Recovering dense models is often achieved using stereovision techniques. Unfortunately, reconstruction artifacts are common and are mainly due to the difficulty to match points on untextured areas of the face between images. In this paper, we propose a robust and fully automatic method to produce realistic dense animation. Our input data are a dense 3D mesh of the talker obtained for one viseme as well as a corpus of stereo sequences of a talker painted with markers that allows the face kinematics to be learned. The main contribution of the paper is to transfer the kinematics learned on a sparse mesh onto the 3D dense mesh, thus allowing dense facial animation. Examples of face animations are provided which prove the reliability of the proposed method.