Abstract
3D face alignment from monocular images is a crucial process in computer vision with applications to face recognition, animation and other areas. However, most algorithms are designed for faces in small to medium poses (below 45 degrees), lacking the ability to align faces in large poses up to 90 degrees. At the same time, many methods are not efficient. The main challenge is that it is time consuming to determine the parameters accurately. In order to address this issue, this paper proposes a novel and efficient end-to-end 3D face alignment framework. We build an efficient and stable network model through Depthwise Separable Convolution and Densely Connected Convolutional, named Mob-DenseNet. Simultaneously, different loss functions are used to constrain 3D parameters based on 3D Morphable Model (3DMM) and 3D vertices. Experiments on the challenging AFLW, AFLW2000-3D databases show that our algorithm significantly improves the accuracy of 3D face alignment. Model parameters and complexity of the proposed method are also reduced significantly.