Better initialization for regression-based face alignment

Regression-based face alignment algorithms predict facial landmarks by iteratively updating an initial shape, and hence are always limited by the initialization. Usually, the initial shape is obtained from the average face or by randomly picking a face from the training set. In this study, we discuss how to improve initialization by studying a neighborhood representation prior, leveraging neighboring faces to obtain a high-quality initial shape. In order to further improve the estimation precision of each facial landmark, we propose a face-like landmark adjustment algorithm to refine the face shape. Extensive experiments demonstrate our algorithm achieves favorable results compared to the state-of-the-art algorithms. Moreover, our algorithm achieves a smaller normalized mean error than the human performance (5.54% vs. 5.6%) on the challenging dataset the Caltech Occluded Faces in the Wild (COFW).