A deep Coarse-to-Fine network for head pose estimation from synthetic data

Various applications of human-computer interaction are based on the estimation of head pose, which is challenging due to different facial appearance, inhomogeneous illumination, partial occlusion, etc. In this paper, we propose a deep neural network following the Coarse-to-Fine strategy to estimate head poses. The scheme includes two branches: Coarse classification phase classifying the input image into four categories, and Fine Regression phase estimating the accurate pose parameters. The two sub-networks are trained jointly. To tackle the problem of insufficient annotated data in training process, we design a rendering pipeline to synthesize realistic head images and generate an annotated dataset with a collection of 310k head poses. The results on benchmark datasets and synthetic dataset validate the effectiveness of our approach, as well as the results on images with diverse illumination, occlusion, and motion blur. Moreover, our method can be easily extended to estimate head poses on depth images.