Unsupervised Lightweight Face 3D Reconstruction From a Single Uncalibrated Image

Reconstruct 3D model from 2D image is an important task in the field of deep learning, which aims to make computers have the ability to perceive the 3D world like human-beings. In this paper, a lightweight method is proposed for 3D face reconstruction from a single image without any supervision. Specifically, our method employs encoder-decoder architectures to extract depth map, light condition, transformation matrix and albedo from input image. According to the principle of rendering, we can obtain the connection between 2D image coordinates and 3D model vertices coordinates, using the above elements, we can get the projection image of the reconstructed model. Then the reconstruction loss can be used to optimize the network parameters. Experiments show that our method surpasses the previous work in reconstruction speed and the size of model.