Robust And Detailed Reconstuction Of Human Face From Single Image

Recently, 3D face reconstruction based on 3D Morphable Model (3DMM) has made great progress. However, 3D face reconstruction based on a single face image is always an ill-posed problem which can not capture detailed textures of faces in most cases. Recent works which recovered the faces with UV maps can not change in shape, and only have some visual details. In this paper, we introduce an end-to-end 3D face reconstruction framework in order to recover detailed 3D face from a single face image. The main idea is utilizing the facial details of the input image to refine the 3DMM-based rough mesh in order to make it more close to the real face in shape. We use the UV displacement map and UV visible mask to recover face details and make the model more robust to occlusions in face images. During the reconstruction, we use the multiple-level loss to make our training process unsupervised. This method essentially improves the problem that detailed faces are hard to be recovered through monocular 3D face reconstruction. We quantitatively evaluate our model on MICC Florence 3D Face dataset and finally get the acceptable results.