Robust facial feature localization with probabilistic constrained local models

Constrained local models (CLMs) recently exhibit superior generic performance in facial feature localization over leading holistic gradient descent methods, such as active appearance models (AAMs). However, due to representing shape variations with a single deterministic PCA (Principal Component Analysis) model, canonical CLMs may suffer inherent drawbacks when applied to faces with large non-rigid shape variations. To solve the problem, this paper presents a probabilistic constrained local model (PCLM) via introducing probabilistic concepts into the canonical CLM framework. The PCLM consists of an ensemble of patch experts estimating likelihood of multiple candidate positions for each facial landmark, and a shape prior model based on mixtures of probabilistic principal component analyzers (MPPCA). The shape and pose parameters of the model are estimated using the expectation-maximization (EM) algorithm within a maximum-likelihood framework. Experimental results prove that the proposed PCLM is capable of dealing with face images of large non-rigid shape variations and noises.