A Lightweight 3D Human Body Reconstruction Method Based on Gradient Scaling and Gaussian Distribution Optimization

In this paper, we propose a lightweight, high-quality and effective method for creating 3D digital humans from monocular videos, achieving fast training (2 - 3 minutes) and real-time rendering. Specifically, we use 3D Gaussian Splatting (3DGS) as the basic framework, referring to the relevant work of GauHuman, and propose a gradient scaling strategy. When calculating the Gaussian growth conditions, we consider the number of pixels covered by the Gaussian in each view. The number of covered pixels is used as a weight to dynamically average the gradients from different views, thus promoting the growth of large Gaussians. At the same time, we propose a Gaussian distribution optimization strategy. By controlling the opacity of the Gaussians, we gradually eliminate the Gaussians that contribute little to the rendering, reducing computational redundancy. Extensive experiments on the ZJU_MoCap dataset show that we ensure the rendering quality under real-time rendering and significant lightweighting.