Face alignment using local hough voting

We present a novel Hough voting-based method to improve the efficiency and accuracy of fiducial points localization, which can be conveniently integrated with any global prior model for final face alignment. Specifically, two or more stable facial components (e.g., eyes) are first localized and fixed as anchor points, based on which a separate local voting map is constructed for each fiducial point using kernel density estimation. The voting map allows us to effectively constrain the search region of fiducial points by exploiting the local spatial constraints imposed by it. In addition, a multioutput ridge regression method is adopted to align the voting map and the response map of local detectors to the ground truth map, and the learned transformations are then exploited to further increases the robustness of the algorithm against various appearance variations. Encouraging experimental results are given on several publicly available face databases.