Joint estimation of Pose and Facial Landmark using boosted parallel random ferns

Abstract

Facial landmark localization is important task in various face applications such as face verification, identification, and tracking. It is a key pre-processing step in achieving high performance in an unconstrained environment, that includes pose variation among others such as occlusion and variation to illumination. In this thesis, the facial landmark is considered to be a linear combination of pose-inspired landmarks, and facial landmark and pose are jointly estimated using the proposed boosted parallel random ferns in a forward stage-wise manner. To cope with displacement in pixel indices when projecting a 3D face on a 2D face image, pose-landmark indexed feature is introduced. The method achieves state-of-the-art performances in both facial landmark localization and pose estimation on the LFW and the MultiPIE datasets.