A Survey for Conventional Regression- and Deep Learning-based Face Alignment Methods

Abstract

Face alignment, as an important part of facial tasks, will affect the final efficiency and accuracy. Face alignment is to locate the exact shape of a detected face bounding box. There are amount of challenges in face alignment because of large poses, occlusions and illuminations in real-world conditions. The approaches to tackle these challenges can be categorized in methods based on regression, which require operators in feature extraction, and methods based on deep learning, in which the feature extraction is data driven. Methods applies regression include Supervised Descent Method and Face Alignment by Coarse-to-Fine Shape Searching. Deep Convolutional Neural Networks, Tasks-Constrained Deep Convolutional Network and Multi-task Cascaded Convolutional Networks apply cascaded CNN and they are representational approaches of deep learning method. This article is devoted to the elaboration and summary of these mainstream methods.