EMA-GS: Improving sparse point cloud rendering with EMA gradient and anchor upsampling

Abstract

The 3D Gaussian Splatting (3D-GS) technique combines 3D Gaussian primitives with differentiable rasterization for real-time high-quality novel view synthesis. However, in sparse regions of the initial point cloud, this often results in blurring and needle-like artifacts owing to the inadequacies of the existing densification criterion. To address this, an innovative approach that utilizes the Exponential Moving Average (EMA) of homodirectional positional gradients as the densification criterion is introduced. Additionally, in the early stages of training, anchors are upsampled near representative locations to infill details into the sparse initial point clouds. Testing on challenging datasets such as Mip-NeRF 360, Tanks and Temples, and DeepBlending, the results demonstrate that the proposed method achieves fine detail recovery without redundant Gaussians, exhibiting superior handling of complex scenes with high-quality reconstruction and without requiring excessive storage. The code will be available upon the acceptance of the article.