GaussianAvatar: Human avatar Gaussian splatting from monocular videos

Abstract

Many application fields including virtual reality and movie production demand reconstructing high-quality digital human avatars from monocular videos and real-time rendering. However, existing neural radiance field (NeRF)-based methods are costly to train and render. In this paper, we propose GaussianAvatar, a novel framework that extends 3D Gaussian to dynamic human scenes, enabling fast training and real-time rendering. The human 3D Gaussian in canonical space is initialized and transformed to posed space using Linear Blend Skinning (LBS), based on pose parameters, to learn the fine details of the human body at a very small computational cost. We design a pose parameter refinement module and a LBS weight optimization module to increase the accuracy of the pose parameter detection in the real dataset and introduce multi-resolution hash coding to accelerate the training speed. Experimental results demonstrate that our method outperforms existing methods in terms of training time, rendering speed, and reconstruction quality.