Multi-manifold modeling for head pose estimation

In this paper, we study the identity-independent head pose estimation problem, in order to handle the appearance variations, we consider the pose data lying on multiple manifolds. We present a novel manifold clustering method to construct multiple manifolds each of which characterizes the underlying subspace of some subjects. We first construct a set of n-simplexes of subjects by using the similarity of pose images. Then, we present a supervised method to obtain a low-dimensional manifold embedding for each n-simplex. Finally, we propose the K-manifold clustering method, integrating manifold embedding and clustering, to make each learned manifold with unique geometric structure. The experimental results on a standard database demonstrate that our method is robust to variations of identities and achieves high pose estimation accuracy.