Paper accepted at CVPR 2026

Our paper “PRIMU: Uncertainty Estimation for Novel Views in Gaussian Splatting from Primitive-Based Representations of Error and Coverage” has been accepted for publication at CVPR 2026.

In this work, Thomas Gottwald, Edgar Heinert, Peter Stehr, Chamuditha Jayanga, and Matthias Rottmann present a method to estimate predictive uncertainty for novel views in Gaussian Splatting (3DGS). Given a camera pose, PRIMU provides indications of whether a 3DGS representation delivers a reliable rendering and depth estimate.

This is particularly relevant for downstream applications such as active view selection, robot navigation, and 3D change detection, where knowing the reliability of rendered views is important for automated decision-making.

The work is a joint collaboration between Universität Osnabrück, Bergische Universität Wuppertal, and Queensland University of Technology (QUT), and was supported by the ERDF project JustScanIt3D, the BMFTR project UnrEAL, and DAAD PPP.