Audit reveals overconfidence in feed-forward 3D reconstruction models

zhenjun_zhao · x · 2026-09-01

This paper audits the per-pixel confidence emitted by feed-forward 3D reconstruction models. It finds that while confidence ranks error well, predicted uncertainty is systematically too low on out-of-distribution data. Models are overconfident on unseen frames, with error increasing as confidence rises. Rescaling helps average magnitude but fails to fix scene-level deviations.

Original post →

More from Research

Research channel →