Localisation-Aware Uncertainty for Pretrained Object Detection

2026-10-01

AI Summary

This study addresses the limitation that existing uncertainty estimation methods for pretrained object detectors under distribution shifts or adversarial attacks typically require retraining and incur high computational costs. To this end, this work proposes a lightweight post-hoc evidential meta-model that keeps the base detector frozen without modifying its architecture or requiring repeated inference. Specifically, it automatically identifies localization-relevant features via saliency-guided curriculum learning and constructs detection-level objectives by integrating localization errors with prediction instability, thereby enabling efficient posterior bounding-box uncertainty estimation grounded in evidential deep learning. Experimental results demonstrate that the proposed approach preserves in-domain detection performance while improving the TP-FP AUROC by up to 22% under adversarial attack scenarios, achieving robust and low-cost uncertainty quantification for object detection.

Abstract

Reliable uncertainty estimation is essential for deploying object detectors when distribution/covariate shift and adversarial attacks may occur. Existing approaches often require detector retraining, architectural modification, or repeated inference, which may be infeasible or incur significant overheads. We introduce a lightweight post-hoc evidential meta-model that learns when object localisations should be considered uncertain while keeping the base detector frozen. Our approach automatically identifies localisation-relevant features and uses saliency-guided modification to construct an increasingly challenging curriculum. Detection-level targets combine localisation error, modification level, and prediction instability to guide an evidential meta-model to estimate uncertainty for each predicted bounding box. Our approach requires no changes to the detector and preserves its original localisation outputs. Across adversarial attacks and evaluated strengths, GRACE improves TP-FP AUROC by 22% relative to the strongest comparator in some cases while maintaining in-distribution detection performance.

Problem

• uncertainty estimation
• object detection
• distribution shift
• adversarial attacks
• localisation

Innovation

• Uncertainty Estimation
• Object Detection
• Evidential Meta-Model
• Curriculum Learning
• Adversarial Robustness