Robust Face Pose Estimation from Insufficient Data

This paper presents a novel method to estimate the pose of human faces from insufficient video data. We represent each pose of a person's face as a connected low-dimensional appearance manifolds which are approximated by affine plane. To construct affine planes, we first sample exemplars from video data and cluster exemplars into each pose. From exemplars, the affine plane is constructed using PCA. However, the sampled exemplars in each specific pose are often not enough for computing the affine plane. This limits the performance of pose estimation. To overcome it, we propose a new sample generation method in constructing pose manifold for on-line face manifold learning. The proposed method was evaluated under several real situations and promising results were obtained.