Cascaded Pose Regression Revisited: Face Alignment in Videos

Abstract

Automated pose estimation is a fundamental task in computer vision. In this paper, we investigate the generic framework of Cascaded Pose Regression (CPR), which demonstrates practical effectiveness in pose estimation on deformable and articulated objects. In particular, we focus on the use of CPR for face alignment by exploring existing techniques and verifying their performances on different public facial datasets. We show that the correct selection of pose-invariant features is critical to encode the geometric arrangement of landmarks and crucial for the overall regressor learnability. Furthermore, by incorporating strategies that are commonly used among the state-of-the-art, we interpret the CPR training procedure as a repeated clustering problem with explicit regressor representation, which is complementary to the original CPR algorithm. In our experiment, the qualitative evaluation of existing alignment techniques demonstrates the success of CPR for facial pose inference that can be conveniently adopted to video detection and tracking applications.