Title: Instance-Level Post Hoc Uncertainty Quantification in Object Detection

Abstract: Object detection serves as a critical safety component in autonomous driving systems, making it imperative to measure the uncertainty associated with bounding-box predictions to ensure operational safety. To meet the demands of real-world deployment, which typically prohibits retraining, we utilize the Laplace approximation for post hoc uncertainty quantification. However, existing approaches face limitations: linearized inference methods are computationally inefficient for instance-level needs due to their reliance on multiple backpropagations, while sampling-based techniques do not strictly adhere to post hoc constraints. To address these challenges, we introduce the Monte-Carlo generalized linearized model (MC-GLM), a framework designed to deliver both instance-level and approximately post hoc uncertainty estimates. A key advantage of MC-GLM is that the sample count required for the Monte Carlo step remains constant, regardless of the number of output instances, thereby enabling effective parallelization. Our evaluation on the nuScenes dataset using the CenterPoint detector confirms the method’s efficacy, demonstrating that the generated uncertainties are of high quality.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC