Trust It or Not: Evidential Uncertainty for Feed-Forward 3D Reconstruction with Trust3R

Geometric foundation models hold promise for unconstrained dense geometry prediction from uncalibrated images. However, current feed-forward designs often produce heuristic confidence scores that lack probabilistic interpretation and fail to indicate where and how much the predicted geometry can be trusted. To address this gap, we present Trust3R, a lightweight evidential uncertainty framework for feed-forward 3D reconstruction. Trust3R combines gated residual mean refinement with a Normal-Inverse-Wishart evidential head, yielding a closed-form multivariate Student-t distribution for per-point geometric uncertainty. This provides probabilistically grounded pointmap uncertainty estimates with moderate inference overhead. We evaluate on diverse indoor and outdoor benchmarks and compare against MASt3R's built-in confidence map, single-pass heteroscedastic regression, MC dropout, and deep ensembles. Experimental results show that Trust3R consistently improves risk--coverage and sparsification, generally improves geometric accuracy, and strengthens uncertainty ranking across benchmarks. On ScanNet++, Trust3R achieves 25% lower AURC and 41% lower AUSE, providing a practical reliability signal for uncertainty-aware weighting in downstream geometry pipelines. Our project page and code are available at https://trust3r-z.github.io.

Computers can now build 3D scenes from ordinary images, which is useful for mapping, robotics, and augmented reality. But these systems can still make mistakes in difficult regions, such as occluded, textureless, shiny, or transparent surfaces. A serious problem is that they may look confident even when their 3D prediction is wrong, so later systems may over-trust bad geometry. We introduce Trust3R, a lightweight method that teaches a one-step 3D reconstruction model to say not only what 3D point it predicts, but also how reliable that prediction is. Instead of relying on a simple confidence score, Trust3R gives each predicted 3D point an uncertainty estimate, meaning a measure of how likely that point may be wrong. It does this efficiently, using only a single run of the model. This allows unreliable points to be down-weighted or removed before they affect downstream 3D processing. Across indoor and outdoor benchmarks, Trust3R better aligns uncertainty with actual 3D errors than existing confidence-based approaches. It also improves reliability in downstream settings such as camera pose estimation and difficult transparent-object scenes. This work helps make image-based 3D reconstruction easier to trust, while the uncertainty should still be treated as guidance rather than a guarantee.