Facial Shape Localization Using Probability Gradient Hints

Zhiheng Niu, Shiguang Shan, Xilin Chen

Abstract: This letter proposes a novel method to localize facial shape represented by a series of facial landmarks. In our method, the problem of facial shape localization is formulated with a Bayesian inference. Specifically, given a face image, the posterior probability of the facial shape is naturally decomposed into two parts: the likelihood function of local textures and the prior constraints of global shape. The former is provided by the landmark detectors, while the latter is evaluated based on the global shape statistics. The global shape is iteratively estimated in the Maximum A Posteriori (MAP) procedure which is derived in a Lucas-Kanade manner over the probability distribution. Intuitively, in each step, the landmarks are driven by the probability gradient and converge towards the positions which maximize the posterior probability. Experiments on two public databases (XM2VTS and BioID) show the effectiveness of the proposed method.