Paper accepted for publication at ECCV 2026

Our joint work with Queensland University of Technology (QUT) in Brisbane entitled “Predictive Photometric Uncertainty in Gaussian Splatting for Novel View Synthesis” by Thomas Gottwald, Peter Stehr, Edgar Heinert, Chamuditha Jayanga, Niko Sünderhauf, and Dimity Miller has been accepted for publication at ECCV 2026.

3D Gaussian Splatting has become a powerful technique for novel view synthesis, but safety-critical applications require reliable estimates of when rendered views can be trusted. In this work, we present a lightweight, plug-and-play framework for pixel-wise, view-dependent predictive uncertainty estimation that operates as a post-processing step and leaves the underlying scene representation unchanged.

We demonstrate the effectiveness of our approach across three downstream tasks: active view selection, scene change detection, and anomaly detection, all without requiring pose information. Our results highlight the potential of predictive uncertainty estimation as a meaningful step towards trustworthy 3D perception.

We would like to thank our collaborators Chamuditha Jayanga, Niko Sünderhauf, and Dimity Miller from QUT for the excellent collaboration, as well as our team at the OSVIA Lab, University of Osnabrück / Bergische Universität Wuppertal for their outstanding contributions.

This work is supported with funds by DAAD PPP, ERDF, and JustScanIt3D. The preprint is available here: https://arxiv.org/abs/2603.22786