Sign-correlation cascaded regression for face alignment

Abstract

Face alignment plays an important role in many applications such as face recognition and face reconstruction. Current regression based approaches can ease the multi-pose face alignment problem, but they fail to deal with the multiple local minima problem directly. To improve the performance of multi-pose facial landmark localization, in this paper we propose a sign correlation supervised descent method (SC-SDM) based on a nonlinear optimization theory. SC-SDM analyses the sign correlation between features and shapes and project both of them into a mutual sign-correlation subspace. By partitioning the whole multi-pose samples into a series of pose-consistent subsets, a group of models are learned from each subset. The experiments using the public multi-pose datasets has validated the partition and proved that SC-SDM can accurately separate samples into pose-consistent subsets, which reveals their latent relationships to pose. The comparison with state-of-the-art methods demonstrates that SC-SDM outperforms them, especially in uncontrolled conditions with various poses.