Overview

We propose a method to reconstructe high precision human face based on multi-view 2D images without camera external parameters, which have a great performance on extream experssions and wrinkles. Considering the limits of face mesh(3DMM), we do not use any face priority mesh, and follow the idea of SFM to design our pipeline. Although in some extream lighting conditions, our method has a great performance as well.

Vision Result

In order to have a better result, we estimate camera external parameters and geometry simultaneously, which achieves a great face alignment result campared with initial value.

Thanks to FaceScape team for providing us with the dataset .

We provide a tool to synthesize face data, which can adjust face id, pose, lighting condition. [link]

The website template was borrowed from Michaël Gharbi.