Multiview Facial Feature Tracking with a Multi-modal Probabilistic Model

Dynamically tracking facial features with large variation of face pose has been shown as one of the most challenging issues in facial feature tracking. The traditional statistical models employ a principal component analysis (PCA) to characterize the statistics of a set of example shapes, however they are restricted to a narrow view due to the global linearity assumption. In this paper, a novel model-based multi-modal probabilistic approach is proposed to capture the complicated relationships among facial features under different face poses by using a mixture of local linear probabilistic PCAs. Based on the probabilistic evaluation of the component Probabilistic PCAs, the proposed method provides an effective way to choose the appropriate local PCA automatically and accurately when face poses are undergoing large variations. Experiment results demonstrate that the proposed method could track the facial features robustly with high accuracy under large variations of pose over time