Efficient and robust integration of face detection and head pose estimation

Most previous studies on head pose estimation assume that the face is already detected and aligned. However, this is not the case in many real-world applications. Automatic head pose estimation typically starts with face detection followed by pose estimation on the detected faces. For computational efficiency, it is desirable if the face detector and the head pose estimator are based on the same image features. In this work, we show that normalized Gabor features, previously proposed for face detection, are also a robust representation for pose estimation. In particular, in the normalized Gabor feature space faces with similar poses are closer than in other feature spaces. Pose estimation with these features using nonlinear regression based on the weighted K nearest neighbors performs better than previously reported approaches on the same database under more complex illumination conditions. Because the features have already been computed for the face detector, the combined autonomous system running on a Quad-core i5 2.66 GHz PC requires only 2.56ms of additional computation to estimate pose for each detected face.