Real-time view-based face alignment using an extended AAM

In this paper, we present some extensions to Active Appearance Model and its extended view-based models are applied for tracking faces through wide angles. The extensions include:(1) a new multiband representation created by filtering the images with a set of orthogonal spatial filters and then applying a non-linear normalization to them, and (2)adding extra landmarks outside the facial region to encapsulate the complete contours of chin, cheeks and eyebrows and incorporate more information. The experimental results demonstrate that our extensions can achieve higher accuracy and better convergence rate, and verify that our extended AAM can be reliably applied for real-time face tracking under large pose variations.