Capturing appearance variation in active appearance models

The paper presents an extension of active appearance models (AAMs) that is better capable of dealing with the large variation in face appearance that is encountered in large multi-person face data sets. Instead of the traditional PCA-based texture model, our extended AAM employs a mixture of probabilistic PCA to describe texture variation, leading to a richer model. The resulting extended AAM can be efficiently fitted to held-out test images using an adapted version of the inverse compositional algorithm: the computational complexity scales linearly with the number of components in the texture mixture. The results of our experiments on three face data sets illustrate the merits of our extended AAM.