基于各向异性可见性场的不确定性驱动 3D 高斯溅写主动映射
摘要
我们提出 Gaussian Splatting Anisotropic Visibility Field (GAVIS)，用于 3DGS 的不确定性量化和主动映射。我们的关键洞察是，来自训练视图之外的区域会导致 3DGS 的预测不可靠。为此，我们引入一种原则且高效的方法，用于量化 3DGS 中的可见性场，定义为每个粒子相对于训练视图的各向异性可见性，并使用球面谐波表示。 resulting visibility field is integrated into a Bayesian Network-based uncertainty-aware 3DGS rasterizer, enabling real-time (200 FPS) uncertainty quantification for synthesized views. Active mapping is further performed within a maximum information gain framework building on this formulation. Extensive experiments across diverse environments demonstrate that GAVIS consistently and significantly outperforms prior approaches in both accuracy and efficiency. Moreover, beyond standalone use, our method can be applied post-hoc to improve the performance of existing approaches.
引用
@article{arxiv.2605.30342,
title = {Uncertainty-driven 3D Gaussian Splatting Active Mapping via Anisotropic Visibility Field},
author = {Shangjie Xue and Jesse Dill and Dhruv Ahuja and Frank Dellaert and Panagiotis Tsiotras and Danfei Xu},
journal= {arXiv preprint arXiv:2605.30342},
year = {2026}
}
备注
Accepted to CVPR 2026. Project page https://gatech-rl2.github.io/GAVIS/