Face Alignment via Multi-Regressors Collaborative Optimization

Face alignment is a fundamental step in facial image analysis. To solve the non-convex optimization problem, most cascade-based regression approaches conventionally utilize a single regressor to cover the entire optimization space. These measures are prone to average conflicting gradient directions, especially when applied to faces in the unconstrained condition with various poses and expressions. In this paper, we present an effective face alignment approach based on the multi-regressors collaborative optimization. The foundation of our method is the cascaded regression (CR) that has recently established itself as one of the most practical and effective frameworks for localizing the facial landmarks. CR is interpreted as a learning-based approach to iteratively optimize an objective function. On this basis, in all iterations, the proposed algorithm further divides the sample space into several clusters, in each of which, samples with similar gradient directions and one separate local regressor are learned. During the prediction stage, the unseen landmarks of a face image are evaluated by a linear combination of estimations from all cluster regressors with different weights. The experimental results demonstrate the advantages of our method on the general unconstrained images.