A method for efficient and robust facial features localization

Abstract

We present a fast and robust algorithm for face alignment. There are three key contributions. The first is the introduction of a new shape indexed feature called multi-resolution wrapped features (MRWF), which is robust to scale and poses variation, and can be calculated very efficiently. The second is a new gradient boosting method based on a mixture re-sampling strategy, which allows the model to resistant to imbalance of training samples. The third contribution is a method for localizing facial feature points of an unknown image in a new iterative manner, which makes the algorithm robust to initial location. Extensive experiments over images with obvious pose, expression and illumination changes have shown the accuracy and efficiency of our method. © Springer International Publishing 2013.