Our paper “PRIMU: Uncertainty Estimation for Novel Views in Gaussian Splatting” has been accepted at CVPR 2026. The method estimates how reliable novel 3DGS renderings are, enabling safer use of neural 3D reconstruction for applications such as robotics, view planning, and change/anomaly detection.
In this extended series of experiments, we investigate how the number of images, orientation and pre-processing affect the quality of 3D reconstructions using Gaussian splatting. Using interactive web demos, we compare different data sets and show how even large numbers of images can be processed efficiently and presented with high visual quality.
On April 11, 2025, the second project meeting took place at the German Mining Museum in Bochum at the invitation of Prof. Michael Prange.
In the summer of 2024, the "Just Scan It 3D" project began, using AI, uncertainty quantification, and anomaly detection to authentically capture 3D scenes and models – exemplified by the Doppelbock of the Deutsches Bergbau-Museum Bochum, which was created using drone footage and 3D Gaussian Splatting.