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

摘要: Abstract Various applications of human-computer interaction are based on the estimation 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 Coarse-to-Fine strategy estimate poses. The scheme includes two branches: Coarse classification phase classifying input image into four categories, and Fine Regression estimating accurate pose parameters. sub-networks trained jointly. To tackle problem insufficient annotated data in training process, design rendering pipeline synthesize realistic images generate an dataset with collection 310k results benchmark datasets synthetic validate effectiveness our approach, as well diverse motion blur. Moreover, method can be easily extended poses depth images.