Face alignment using two-stage cascaded pose regression and mirror error correction

Abstract

Recently, a series of cascaded pose regression based facial landmark localization methods under occlusion have been proposed. However, partial occlusions and pose variations will break the entire structure of the face which poses obstacles to global regression. Moreover, there lack techniques to evaluate the reliability of the regression results during the regression process. In this paper, we propose a Two-Stage Cascaded Pose Regression(TSCPR) for facial landmark localization under occlusion. In the first stage, a global cascaded pose regression with robust initialization is performed to get localization results for the original face and its mirror image. The localization difference between the original image and the mirror image is used to determine whether the localization of each landmark is reliable, while unreliable localization can be adjusted. In the second stage, the global results are divided into multiple parts, which are further refined by local regressions. Finally, multiple refined local results are rated and adjusted to get the final output. We evaluated the proposed method on widely used datasets COFW, LFPW, HELEN, 300-W and Menpo-Semifrontal. The experimental results show that the proposed method can outperform the state-of-the-arts.