Object detector uncertainty improves localization without retraining
Localisation-Aware Uncertainty for Pretrained Object Detection
Computer Vision and Pattern Recognition
Summary
Object detectors need to know when their predictions might be wrong, especially under difficult conditions like attacks or unfamiliar situations. The paper presents a lightweight add-on that learns to estimate how uncertain each detected object’s position is, without changing the original detector or needing extra runs. It uses intelligent tweaks of features related to object location and trains a model to predict uncertainty based on errors and instability in detection. This method keeps the detection accuracy steady while better identifying unreliable boxes during attacks or shifts.
What this means in practice
- •For autonomous vehicle engineers: Improve object detector reliability by flagging uncertain object locations without altering existing detection systems.
- •For security system developers: Detect when object localization from cameras is unreliable under adversarial attacks to maintain safer monitoring.
Authors
Charmaine Barker, Daniel Bethell, Simos Gerasimou
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.