A robust composite metric for head pose tracking using an accurate face model

We present a method for full-motion recovery of the head pose from a monocular video input based on an accurate head model and textures. We build our face model using a distribution of high resolution 3D face scans. The cost of computations makes us select parts of this full model. To address the difficult task of initializing the model position and tracking its motion, we use a composite metric using face texture samples from three different face databases, following a positive face detection. The algorithm is subsequently able to track the head in a wide pose range with great accuracy. We also provide test video sequences with an independent accurate ground truth (with estimated RMS error of 0.1 degrees).