Facial feature points detecting based on Gaussian Mixture Models

• We use Gaussian Mixture Models to approximate the confidence maps analytically.
• We also use Gaussian Mixture Models to model the distribution of face shape.
• The posterior is maximized by iteratively maximizing a lower bound of it.
• High-order geometry constraint within facial features are considered.
• Model based Hough Voting scheme is proposed to estimate an initial face shape.

Detecting predefined facial feature points in a human face image is a well studied problem. Despite the impressive achievements that have been made, it is still open under unconstrained environments, with variations of illumination, expression, head pose, as well as partial occlusions. This paper proposed a novel method to locate facial feature points under the variations mentioned above. Support vector machines with probability outputs are trained to provide the observation probability of each facial feature point. The observation probabilities, as well as the distribution of face shape which serves as the prior to constrain the relative position of the facial feature points, are both approximated with Gaussian Mixture Models. The problem is solved by maximizing the posterior which combines the prior and observation probability within the framework of Bayesian Inference. An optimization algorithm is developed to maximize the posterior by iteratively maximizing the lower bound of it. The proposed method preserves the high-order geometric constraint within facial feature points. With a simple initialization method of Model based Hough Voting, the method shows competitive detecting rate and locating accuracy on the LFPW and LFW datasets, compared to the methods of state-of-the-art.