Block-regressed face alignment algorithm based on supervised descent method

In recent years, the accuracy of face alignment has been improved a lot. However, the landmarks of each component in face are trained together in most algorithms, which ignore their dedicated characteristic and limit the further accuracy improvement. In this paper, we propose a new approach to improve the localization performance of facial landmarks by taking each component as independent task instead of the whole face. Namely, independent regressors are learned for each component using the gradient descent method. If only considering independent regressors for component, the inherent correlation between the components may be neglected. This paper proposed a strategy to effectively combine the results of the whole face regression and the independent components regressions. In this way, the effect of holistic and independent results are all taken into consideration, which can further enhance the alignment accuracy. A large number of experiments show that our method is better than the single loss function in both detection accuracy and reliability.