Bidirectional warping of Active Appearance Model

Active Appearance Model (AAM) is a commonly used method for facial image analysis with applications in face identification, tracking, and expression recognition. This paper proposes a new approach for AAM fitting. Our approach is called bidirectional warping and is based on image alignment which simultaneously warps both input image and the appearance model when fitting AAM into the input image. Our bidirectional warping technique makes the fitting algorithm less sensitive to initial condition and the model can fit into faces that are never seen or included in the training set. Our experimental results show that our approach outperforms state-of-the-art AAM fitting techniques particularly when the model is far from ground truth.