Uncertainty-driven 3D Gaussian Splatting Active Mapping via Anisotropic Visibility Field

Problem
This paper addresses the gap in uncertainty quantification and active mapping in 3D Gaussian Splatting (3DGS) frameworks, particularly focusing on the limitations of existing methods in handling regions that are not visible from training views. The authors highlight that predictions in these unseen regions are often unreliable, which can significantly impact the performance of 3D mapping systems. This work is presented as a preprint and has not yet undergone peer review.

Method
The authors introduce the Gaussian Splatting Anisotropic Visibility Field (GAVIS), which quantifies the visibility of each particle in the 3DGS using a novel anisotropic visibility field representation. This visibility field is defined in relation to the training views and is modeled using spherical harmonics, allowing for a compact and efficient representation. The GAVIS framework integrates this visibility field into a Bayesian Network-based rasterizer that enables real-time uncertainty quantification at 200 frames per second (FPS). The active mapping component is formulated within a maximum information gain framework, leveraging the visibility field to optimize the mapping process. The training compute details are not disclosed, but the method emphasizes efficiency and real-time performance.

Results
GAVIS demonstrates significant improvements over baseline methods in both accuracy and efficiency across various environments. The paper reports that GAVIS outperforms prior approaches by a notable margin, although specific numerical results and effect sizes are not detailed in the summary provided. The authors conduct extensive experiments to validate their claims, indicating that GAVIS not only excels in standalone applications but can also enhance the performance of existing 3DGS methods when applied post-hoc.

Limitations
The authors acknowledge that their method may still struggle in highly dynamic environments where visibility can change rapidly. Additionally, the reliance on training views may limit the applicability of GAVIS in scenarios with sparse or non-representative training data. The paper does not discuss potential computational overheads associated with the Bayesian Network integration or the scalability of the method to larger datasets.

Why it matters
The implications of this work are significant for the fields of computer vision and robotics, particularly in applications requiring robust 3D mapping under uncertainty. By providing a principled approach to visibility quantification, GAVIS enhances the reliability of predictions in unseen regions, which is crucial for autonomous navigation and environment interaction. The ability to apply GAVIS post-hoc to existing methods suggests a pathway for improving legacy systems without complete overhauls, thereby facilitating broader adoption of uncertainty-aware mapping techniques.

Authors: Shangjie Xue, Jesse Dill, Dhruv Ahuja, Frank Dellaert, Panagiotis Tsiotras, Danfei Xu
Source: arXiv:2605.30342
URL: https://arxiv.org/abs/2605.30342v1