How frontal is a face? Quantitative estimation of face pose based on CNN and geometric projection

Abstract

Face pose estimation has been widely used into various applications of human–computer interaction; however, it is yet a challenging work due to illumination, background, face orientations, appearance visibility, etc. In this paper, a novel coarse-to-fine method of face pose quantitative estimation based on convolutional neural networks (CNN) and geometric projection is proposed. In coarse classification, CNN is applied to classify the input image into a specific category and obtain a relevant weight. After that, geometric projections of 3D face landmarks projected into three planes, x–y, x–z and y–z, of 3D coordinate systems are used to perform the fine estimation of face pose, which can get the offset angles of the face in the three directions of roll, yaw, and pitch. Finally, the final score of face pose is obtained by combining the results of two stages. Experiments on standard datasets show that the proposed method can get better results than some competitive algorithms, which proves the effectiveness of the proposed method.