GAVIS: Real-Time Uncertainty Quantification for 3D Gaussian Splatting via Anisotropic Visibility Fields
GAVIS introduces a principled uncertainty quantification framework for 3D Gaussian Splatting (3DGS) by modeling the anisotropic visibility of each Gaussian particle relative to training views using spherical harmonics. Unseen regions in 3DGS produce unreliable predictions, and GAVIS addresses this by embedding the resulting visibility field into a Bayesian Network-based rasterizer that runs at 200 FPS. The framework extends naturally to active mapping via a maximum information gain strategy, outperforming prior methods in both accuracy and efficiency across diverse environments. Notably, GAVIS can also be applied post-hoc to augment existing 3DGS pipelines without requiring retraining from scratch.