Multi-template supervised descent method for face alignment

Abstract

Supervised Descent Method (SDM) is a highly efficient and accurate approach for facial landmark locating and face alignment. In the training phase, it learns a sequence of descent directions to minimize the difference between the estimated shape and the ground truth in feature space. Then in the testing phase, it utilizes these descent directions to predict shape increment iteratively. However, when the facial expression or direction changes too much, the general SDM cannot obtain good performance due to the large variations between the initial shape and the target shape. In this paper, we propose a multi-template SDM (MtSDM) which can maintain high accuracy on training data and meanwhile improve the accuracy on testing data. Instead of only one model is constructed in the training phase, multiple different models are constructed by repeatedly inputting the images which have large variations on expressions or head poses. And in the testing phase, the distances between some specific landmarks are calculated to select an optimal model to update the point location. The experimental results show that our proposed method can improve the performance of traditional SDM and performs better than several existing state-of-the-art methods.