Pixel-GS: Density Control with Pixel-Aware Gradient for 3D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) has demonstrated impressive novel view synthesis results and advancing real-time rendering performance. However, the effectiveness of 3DGS heavily relies on the quality of the initial point cloud, as poor initialization can result in blurring and needle-like artifacts. This issue is mainly due to the point cloud growth condition, which only considers the average gradient magnitude of points from observable views, thereby failing to grow for large Gaussians that are observable from many viewpoints while many of them are only covered in the boundaries. To address this, we introduce Pixel-GS to take the area covered by the Gaussian in each view into account during the computation of the growth condition. The covered area is employed to adaptively weigh the gradients from different views, thereby facilitating the growth of large Gaussians. Consequently, Gaussians within the regions with insufficient initializing points can grow more effectively, leading to a more accurate and detailed reconstruction. Besides, we propose a simple yet effective strategy to suppress floaters near the camera by scaling the gradient field according to the distance to the camera. Extensive qualitative and quantitative experiments validate that our method achieves state-of-the-art rendering quality while maintaining real-time rendering, on challenging datasets such as Mip-NeRF 360 and Tanks  & Temples. Code and demo are available at: https://pixelgs.github.io.