Warping Non-Frontal Faces to Frontal View

Abstract

Most face recognition systems are optimized for frontal face recordings and even relatively small deviations from the frontal pose (above 10-15 degrees) can cause severe degradation of recognition performance. Therefore, it can be useful to try to compensate for the pose variation and transform the image to a frontal view. It is to be expected that this can improve recognition results. This paper provides a simple warping methid that transforms a face under pose to frontal view using a single fixed 3D model as an approximation of all input faces.