Probabilistic Hierarchical Face Model for Feature Localization

Facial feature localization is an important research area in both computer vision and pattern analysis. We present in this paper a hierarchical face model. It unifies both the global and local face appearance together with geometric information for precise feature localization. Face detection is used to obtain the face rough position and scale in the image. A coarse global AAM (active appearance model) is then applied to obtain the initial position of each part. Local detailed part-level AAM models (like the eyes, mouth) are used to refine the localization. This coarse-to-fine scheme constraints the search in the correct space which makes the localization more accurate. In the part-level localization, we consider the geometric relationship of the local parts and the global face model. The whole procedure is formulated into a maximum-a-posteriori (MAP) framework. Experiments demonstrate the effectiveness of our approach