Supervised descent method based on appearance and shape for face alignment

Abstract

Regression approaches have been recently shown to achieve state-of-the-art performance for face alignment. As a general optimization problem, face alignment is approximately solved by learning a series of mapping functions from local appearance to the coordinates increment of the pixels to detect. There have been extensive studies and continuous improvements have been made in recent years. However, most of the existing methods only rely on the current facial texture in every iteration. It is unreliable to only rely on local appearance information when facial landmarks are partially occluded in unconstrained scenarios. In this paper, a modified supervised descent method is proposed to settle the issue, utilizing both appearance and shape information in learning regression functions. Hence, we call it asSDM. The major contribution of our proposed method is to jointly capture shape and local appearance in cascade regression framework. We evaluate the performance of the proposed method on different data sets and the experimental results on benchmark databases demonstrate that our proposed method outperforms previous work for facial landmark detection.