Calibrated Uncertainty Maps for Radiative Gaussian Splatting in Sparse-View CT
Researchers examine whether uncertainty maps reliably indicate where a sparse-view CT reconstruction is erroneous, and find a discrepancy between whole-volume scoring and localization of errors within the imaged object. They derive analytic, clamp-aware moments to produce calibrated uncertainty estimates for radiative Gaussian splatting reconstructions. The work suggests standard variance-based uncertainty is not equivalent to an error map in this setting.