A fully automatic framework for building 3D animated avatars

The face animation process still requires an important amount of manual human interaction involving highly qualified artists with strong knowledge in facial anatomy. In this paper, we introduce a framework for automatic construction of 3D animated avatars. The proposed approach allows for animating an inanimate facial 3D mesh. The adopted representation is based on the MPEG-4 standard. First, the Supervised Descent Method (SDM) is employed for detecting automatically the MPEG-4 feature points starting from 2D images consisting of projections of 3D avatars. The originality of the approach consists of the training procedure proposed. Thus, we specifically train a SDM detector for each part of the face, i.e. mouth, eyes, nose, chin, cheeks, etc. Skinning operations are performed by jointly using the MPEG-4 FAPs and the feature point based deformation method. The combination of these three methods allows for building animated avatar with a large variety of shapes, i.e. from realistic human faces to cartoons.