HiFace: Hybrid Task Learning for Face Reconstruction from Single Image

Abstract

The task of 3D face reconstruction in the WCPA challenge requires a monocular image as input and outputs 3D face geometry, which has been a prevalent field for decades. Considerable works have been published, in which PerspNet significantly outperforms the other methods under perspective projection. However, as the UV coordinates distribute unevenly, the UV mapping process introduces inevitable precision degradation in dense regions of reconstructed 3D faces. Thus, we design a vertex refinement module to overcome the precision degradation. We also design a multi-task learning module to enhance 3D features. By carefully designing and organizing the vertex refinement module and the multi-task learning module, we propose a hybrid task learning based 3D face reconstruction method called HiFace. Our HiFace achieves the 2nd place in the final official ranking of the ECCV 2022 WCPA Challenge, which demonstrates the superiority of our HiFace.