Facial Landmark Detection using Gaussian Guided Regression Network

Facial landmark detection is important prior information for other face alignment problems such as head pose estimation, facial emotion expression, and face modeling. Among the typical facial landmark detection algorithms, the TREE [1] algorithm which uses the cascaded regression method can detect facial landmark faster than the other algorithms using small training data. In this paper, we extract the Gaussian guided landmark map from the TREE and predict residual facial landmarks via landmark regression network. By using the Gaussian landmark feature map as prior information, the regression network effectively guides the direction of the refined landmarks on the 300W test dataset.