Face Reconstruction With Full Realistic Texture Recovery From A Single Image

3D face reconstruction from a single image often fails in faces with large selfocclusions and variant of illumination.In order to deal with these problems,we present a novel solution to reconstruct 3D face model with full realistic texture in this paper.It not only obtains a promising result in fitting geometry,but also completes the task of recovering full realistic texture.The proposed method can be divided into two parts,geometry reconstruction and texture completing.In the former part,we propose a straightforward method that could reconstruct 3D facial geometry in an end-to-end manner.To achieve this,we design a 2D representation called UV position map which records the 3D shape of a complete face in UV space,then train a simple Convolutional Neural Network to regress it from a single 2D image.Our method does not rely on any prior face model,and can reconstruct full facial geometry along with semantic meaning.Meanwhile,our network is very light-weighted and spends only 9.8ms to process an image,which is extremely faster than previous works.Experiments on multiple challenging datasets show that our method on surpasses other state-of-the-art methods on face geometry reconstruction by a large margin.In the latter part,the illuminated appearance is replaced with mean appearance firstly to relieve the variations and dependence of training dataset,then the occluded texture is completed with a Generative Adversarial Network(GAN).The usage of the information of the visible texture and the high generation ability of GAN helps enhance the robustness of texture completion.Experiments show that our method is invariant to light conditions.Furthermore,the results from these two parts are not only robust to poses and illuminations,but also reserve the idenetiy information of input face,which is significantly useful in many other tasks like face swapping,reenactment and recognition.