Head pose estimation based on extended non-negative matrix factorization

One popular solution to head pose estimation is to formulate it as a pattern classification problem, and treat the holistic facial appearance as the input to classifiers. However, since the face appearance contains all kinds of information, the variation caused by other factors such as identity, expression and lighting may be larger than that caused by different head poses. Thus, the key challenge of these appearance based methods lies in constructing a feature subspace that could successfully recovers head pose while ignoring other sources of image variation. In this paper, following the intuition of combining parts to form a whole face, non-negative matrix factorization (NMF) is extended to learn a localized non-overlapping subspace representation for head pose estimation. To emphasize the appearance variation in head poses, one individual extended NMF subspace is learned for each pose. The head pose of a given face image is then estimated based on its reconstruction error after being projected into the learned pose subspaces. Experiments based on benchmark face database demonstrate the efficiency of the proposed method.