2D projective transformation based active shape model for facial feature location

Active shape model statistically represents a shape by a set of well-defined landmark points and can model object variations using principal component analysis. However, the face modeling becomes degenerate when the test images are taken from different viewpoints other than frontal. In this paper, we present a 2D projective transformation based active shape model to address this problem. First, a 2D projective transformation based alignment algorithm is proposed to model pose variations. Then, projective local profiles are calculated, which make the intensity profiles be scale independent. We evaluate our approach on two different datasets containing both frontal facial images and images with a large range of variations in pose and expression. Experimental results demonstrate the efficiency and effectiveness of the proposed approach.