Shape Augmented Regression for 3D Face Alignment

Abstract

2D face alignment has been an active topic and is becoming mature for real applications. However, when large head pose exists, 2D annotated points lose geometric correspondence with respect to actual 3D location. In addition, local appearance varies more dramatically when subjects are with large pose or under various illuminations. 3D face alignment from 2D images is a promising solution to tackle this problem. 3D face alignment aims to estimate the 3D face shape which is consistent across all poses. In this paper, we propose a novel 3D face alignment method. This method consists of two steps. First, we perform 2D landmark detection based on the shape augmented regression. Second, we estimate the 3D shape using the detected 2D landmarks and 3D deformable model. Experimental results on benchmark database demonstrate its preferable performances.