Query2Uncertainty: Robust Uncertainty Quantification and Calibration for 3D Object Detection under Distribution Shift

Reliable uncertainty estimation is critical for safe autonomous systems, but modern 3D object detectors remain poorly calibrated—especially when the test data differs from training data (distribution shift). Standard post-hoc calibration methods work well on in-distribution data but fail under shift. Now, a team from RWTH Aachen, ING e.V., and KU Leuven present Query2Uncertainty, a new approach that leverages the latent query features of DETR-style 3D detectors (like DETR and its variants). By fitting a density estimator on these compact, location- and class-aware query features, the method adapts model confidences to shifted scenarios. It jointly recalibrates both classification confidence and bounding box regression uncertainty, addressing a key limitation in prior work.

The method was tested on both multi-view camera-based and LiDAR-based detectors, consistently outperforming standard post-hoc methods (e.g., temperature scaling, isotonic regression) in both in-distribution and distribution-shifted settings. The paper, accepted at CVPR 2026, also releases code for reproducibility. This work directly improves the reliability of perception systems in autonomous driving, especially when encountering novel environments or sensor noise. It offers a practical, plug-and-play solution for any DETR-style detector, requiring only the query features already computed during inference.