Abstract
Head pose estimation plays an important role in face recognition. However, it faces vast challenges on account of the initialization, facial feature points’ location accuracy and so on. Inspired by the observation that head pose angles change smoothly and continuously, we present a method based on a robust convolutional neural network for head pose estimation. The proposed network architecture consists of three levels and each level has three convolutional neural networks. The first level is a global one; it predicts the head pose quickly as a preliminary estimation. The following two levels are local ones; they refine the estimation achieved from the previous level step by step. Higher and higher resolution image with different input regions are taken as input in our network. At last, a multi-level regression is employed to combine the estimations from each level. The whole process is conducted in a cascade way to improve the head pose estimation performance directly with three angles together. We perform large experiments on nine challenging benchmark datasets. The experimental results demonstrate that our method performs better than the compared methods.