RF-GS: Reducing Floaters with High-order Regularization for 3D Gaussian Splatting

Recently, 3D Gaussian Splatting (3DGS) has attracted great attention in novel view synthesis for its ability to enhance real-time rendering performance. However, it often produces artifacts in outdoor scenes and strong-light indoor environments. From the visualization results, we observe that this issue stems primarily from the lack of an explicit opacity constraint in 3DGS, which hinders the optimization of abnormally high-opacity Gaussian points (i.e., floaters). These floaters occlude distant Gaussian points, resulting in blurred artifacts in the rendered results. To address these challenges, we propose RF-GS (Reducing Floaters for GS), which utilizes an opacity-aware optimization strategy to mitigate artifacts caused by highly opaque abnormal Gaussian points. Additionally, we adopt a coarse-to-fine pruning strategy to remove redundant points and floaters to improve rendering quality. Extensive experiments demonstrate that our method achieves superior rendering quality and reduces storage costs. Specifically, it outperforms all compared baselines on three challenging datasets including Mip-NeRF 360, Tanks & Temples and the Real-World dataset.