GAVIS: Uncertainty-Driven 3D Gaussian Splatting Active Mapping via Anisotropic Visibility Fields

Forum topic · 小凯 · 2026-06-01

Summary

GAVIS (Gaussian Splatting Anisotropic Visibility Field) is a new framework for uncertainty quantification and active mapping in 3D Gaussian Splatting (3DGS). The key insight is that regions not seen by training views produce unreliable predictions in 3DGS. The authors introduce a principled and efficient method to quantify the visibility field in 3DGS, defined as the anisotropic visibility of each particle relative to training views and represented with spherical harmonics. This visibility field is integrated into a Bayesian-network-based uncertainty-aware 3DGS rasterizer, enabling real-time uncertainty quantification of synthesized views at 200 FPS. Active mapping is then built on top of this via a maximum information gain framework. Extensive experiments across diverse environments show GAVIS consistently and significantly outperforms prior methods in both accuracy and efficiency. Additionally, beyond standalone use, the method can be applied post-hoc to improve existing approaches. Paper: arXiv 2605.30342, by Shangjie Xue, Jesse Dill, Dhruv Ahuja, Frank Dellaert, Panagiotis Tsiotras, and Danfei Xu.

Paper Overview

Field: Computer Vision Authors: Shangjie Xue, Jesse Dill, Dhruv Ahuja, Frank Dellaert, Panagiotis Tsiotras, Danfei Xu arXiv: 2605.30342

Abstract

We present GAVIS (Gaussian Splatting Anisotropic Visibility Field), a new framework for uncertainty quantification and active mapping in 3DGS. The core insight is that regions unseen by training views yield unreliable predictions in 3DGS. To address this, we introduce a principled and efficient method to quantify the visibility field in 3DGS, defined as the anisotropic visibility of each particle relative to training views, represented with spherical harmonics. The resulting visibility field is integrated into a Bayesian-network-based uncertainty-aware 3DGS rasterizer, enabling real-time (200 FPS) uncertainty quantification of synthesized views. Active mapping is achieved on top of this via a maximum information gain framework. Extensive experiments across diverse environments show that GAVIS consistently and significantly outperforms prior methods in both accuracy and efficiency. Moreover, beyond standalone use, our method can be applied post-hoc to boost the performance of existing approaches.

Key Contributions

A principled, efficient quantification of the 3DGS visibility field using anisotropic particle visibility and spherical harmonics
A Bayesian-network-based uncertainty-aware 3DGS rasterizer achieving 200 FPS uncertainty quantification
Active mapping driven by maximum information gain
Post-hoc applicability to improve existing methods

--- *Auto-collected on 2026-06-01*