Face Landmark Calibration Based on 3D Reconstruction and Deep Learning

Abstract

Face alignment is an important research field in machine vision, which have applied in face recognition, expression recognition, pose estimation, and face synthesis. However, many algorithms are design for faces in small poses and good lighting, they do not perform well in the wild. In this paper, we propose a face alignment algorithm to improve the accuracy of face landmark. This algorithm begins with a network to locate face landmark in general and employs a face normalization to reduce disturbing redundant information. Then a 3D facial reconstruction is introduced to get exact camera matrix. Euler angle based on 3D reconstruction eliminates the effect of pose for face landmark. Experiments on the IBUG dataset show that the algorithm achieves improvement over state-of-the-art algorithms.