Studies On Facial Landmark Localization Using Explicit Shape Regression

Robustly localizing facial landmarks often plays a very important role in many multimedia and vision applications, including face tracking, pose estimation and expression analysis. However, facial landmarks location is still a challenging problem, due to complex environment in the wild, like illumination, pose or viewpoint and partial occlusions in the wild. In order to overcome those challenges, geometrical constraints are always vital for facial analysis. The classical methods explicitly build parametric models from face shapes and through continuously iterative optimization. However, viewpoint changes introduce nonlinear factors, active shape model is not robust to images with viewpoint changes and the optimize process is prone to local minima. In recent years, cascade shape regression and deep learning are popular in face alignment. Researchers develop implicit shape regression, generating mappings from texture to shape can process viewpoint changes in a certain extent. However, most recently proposed regression-based methods highly depend on availability of training examples and regression methods. Deep learning methods give rise to over-fitting on lack of training data.The solution proposed in this paper extract facial feature points and utilize shape priors to locate points in a shape-to-gradient regression framework. Human faces have common geometrical structures invariant to viewpoint changes. The algorithm introduces a novel projective invariant, named characteristic number (CN), to explicitly characterize the intrinsic geometries of facial points shared by human faces. By further developing a shape-to-gradient regression framework, the computation of our model can be successfully addressed by learning the descent directions using point-CN pairs without the need of large collections for appearance training. Extensive experiments on challenging benchmark data sets like LFW, LFPW and Helen demonstrate the effectiveness of our proposed detector against some state-of-the-art approaches to a certain extent. Moreover, we build a Multi-viewpoint data set which is collected in the indoor environment, where each face has 15 well-defined viewpoints (poses) in order to quantitatively analyze the effects of different poses to localization methods. Extensive experiments show that our shape regression is robust on training set.