Realistic face animation generation from videos

TL;DR

This paper reviews recent deep learning methods for 3D face reconstruction and alignment from videos, highlighting their advantages and potential improvements for more realistic face animation generation.

Contribution

It introduces and analyzes three state-of-the-art methods, proposing enhancements to improve accuracy and speed in 3D face reconstruction.

Findings

Analysis of three leading 3D face reconstruction methods

Proposed improvements to PRN for better performance

Insights into the limitations of pre-defined face templates

Abstract

3D face reconstruction and face alignment are two fundamental and highly related topics in computer vision. Recently, some works start to use deep learning models to estimate the 3DMM coefficients to reconstruct 3D face geometry. However, the performance is restricted due to the limitation of the pre-defined face templates. To address this problem, some end-to-end methods, which can completely bypass the calculation of 3DMM coefficients, are proposed and attract much attention. In this report, we introduce and analyse three state-of-the-art methods in 3D face reconstruction and face alignment. Some potential improvement on PRN are proposed to further enhance its accuracy and speed.