A brief review of 3D face reconstruction methods for face-related product design

Abstract

3D face reconstruction is highly important in the ergonomics study of 3D face, especially in terms of designing face-related products. With the development of machine vision and deep learning, it becomes feasible to reconstruct the 3D face from a single image, which can make it practical to obtain a large scale data of 3D face shape instead of using the 3D scanning technology. The 3D face reconstruction methods, to recover the 3D facial geometry under unconstrained situations from 2D images, are roughly classified into two categories, namely (1) 3D Morphable Model (3DMM) fitting based method and (2) End-to-end deep convolutional neural network (CNN) based method. The 3DMM as a general face representation is introduced emphatically and two kinds of 3DMM fitting based methods are introduced when improving the 3DMM modeling mechanism. Four representative CNN based methods are compared when regressing from pixels of face image to the 3D face coordinates in different gird-like data structures. Finally, six common face datasets largely used in the training and testing are listed.