Feature Fusion for Facial Landmark Point Location

Abstract

Size of regions and discrimination of features are important to local approaches for facial landmark point location. After size of regions is determined, importance of features is obvious. Three features are considered in the paper, i.e. scale invariant feature transform (SIFT) features, local binary pattern (LBP) features, and Gabor wavelet features. Three logistic regressors are trained by using them respectively, and then they are used to testify the effect of these features on classification performance. On this basis, three fusion features are considered, i.e. SIFT features and Gabor wavelet features, SIFT features and LBP features, and Gabor wavelet features and LBP features. Three logistic regressors are trained by using them, respectively. They will be used to testify the effect of these fusion features on classification performance compared with single features. It can be noted from experiments that fusion features are more discriminative than single features, and that all of the features in the fusion features cooperate at classification stage, instead of single features.