Improving face tracking with 2D template warping

Tracking the face of a computer user as he looks at various parts of the screen is a fundamental tool for a variety of perceptual user interface applications. The authors have developed a simple but surprisingly robust tracking algorithm based on template matching and applied it successfully. This paper describes extensions to that algorithm, which improves performance at large facial rotation angles. The method is based on pre-distorting the single training template using 2D image transformations to simulate 3D facial rotations. The method avoids many of the problems associated with using a complex 3D head model. It is robust to variations in the environment and well-suited to use in practical applications in typical computing environments.