Shape Augmented Regression Method for Face Alignment

Yue Wu, Qiang Ji; Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2015, pp. 26-32

There have been tremendous improvements of the face alignment algorithms, among which the regression framework becomes the most popular one recently. The regression based works start from an initial face shape, and they learn regression models to predict the face shape updates based on the shape-indexed local appearance features. However, most of the regression methods ignore the fact that the regression function should directly rely on the current shape (e.g. regression function for frontal face should be different from that for the left profile face). To utilize this information and improve over the existing regression based methods, we propose the shape augmented regression method for face alignment where the regression function would automatically change for different face shapes. We evaluated the performance of the proposed method on both the general "in-the-wild" database and the 300 Video in the Wild (300-VW) challenge data set. The results show that the proposed method outperforms the state-of-the-art works.