Kweon, MinseongCheng, KaiChen, XuejinPark, JinsunCeylan, DuyguLi, Tzu-Mao2025-05-092025-05-092025978-3-03868-268-41017-4656https://doi.org/10.2312/egs.20251042https://diglib.eg.org/handle/10.2312/egs202510423D Gaussian Splatting (3DGS) efficiently renders 3D spaces by adaptively densifying anisotropic Gaussians from initial points. However, in complex scenes such as city-scale environments, large Gaussians often overlap with high-frequency regions rich in edges and fine details. In these areas, conflicting per-pixel gradient directions cause gradient cancellation, reducing the overall gradient magnitude and potentially causing Gaussians to remain trapped in suboptimal positions even after densification. To address this, we propose NoiseGS, a novel approach that integrates randomized noise injection into 3DGS, guiding suboptimal Gaussians selected for densification toward more optimal positions. In addition, to mitigate the instability caused by oversized Gaussians, we introduce an ℓp-penalization on the scale of Gaussians. Our method integrates seamlessly with existing heuristicbased optimization and demonstrates strong generalization in reconstructing complex scenes such as MatrixCity and Building.Attribution 4.0 International LicenseCCS Concepts: Computing methodologies → Point-based models; Computer vision; Machine learning approachesComputing methodologies → Pointbased modelsComputer visionMachine learning approaches10.2312/egs.202510424 pages