Stacked Hourglass Network for Robust Facial Landmark Localisation

With the increasing number of public available training data for face alignment, the regression-based methods attracted much attention and have become the dominant methods to solve this problem. There are two main factors, the variance of the regression target and the capacity of the regression model, affecting the performance of the regression task. In this paper, we present a Stacked Hourglass Network for robust facial landmark localisation. We first adopt a supervised face transformation to remove the translation, scale and rotation variation of each face, in order to reduce the variance of the regression target. Then we employ a deep convolutional neural network named Stacked Hourglass Network to increase the capacity of the regression model. To better evaluate the proposed method, we reimplement two popular cascade shape regression models, SDM and LBF, for comparison. Extensive experiments on four challenging datasets prove the effectiveness of the proposed method.