Anomaly-Informed Confidence Calibration for Vision-Based Safety Prediction

2026-05-20

AI Summary

Existing vision-based safety predictors exhibit overconfidence under distribution shifts, and conventional anomaly detection methods often overlook dynamic anomalies, posing significant safety risks. To address this, this work proposes an online confidence calibration method that requires no retraining. It uniquely integrates perceptual reconstruction error with cognitive uncertainty derived from a world model’s dynamics, further incorporating control-flow statistics. By leveraging lightweight temperature scaling and test-time augmentation, the approach dynamically adjusts prediction confidence during deployment. Evaluated on a real-world DonkeyCar platform across four unseen anomalous scenarios, the method reduces the average Expected Calibration Error from 0.184 to 0.116—outperforming the best baseline by 37%—while preserving the original safety predictor’s architecture and significantly enhancing robustness.

Abstract

Reliable confidence estimates are important for safely deploying vision-based controllers in autonomous racing, where safety predictions must be derived from camera images, yet modern predictors become dangerously overconfident under test-time distribution shifts. We identify a critical perception-dynamics gap in existing anomaly signals: widely used scores, such as autoencoder reconstruction error, capture visual corruptions but miss dynamics anomalies (e.g., actuation bias, latency), where images remain plausible while the trajectory degrades. To address this, we propose an Anomaly-Informed Online Calibration approach that, without retraining any model component, fuses two complementary anomaly scores extracted from a world model: a perceptual score from reconstruction error and a dynamics score from epistemic uncertainty and control-stream statistics. Based on these fused scores, a lightweight temperature-scaling calibrator leverages test-time augmentation to selectively reduce overconfidence under shift while preserving nominal-condition performance. Experiments on a physical DonkeyCar under four real-world anomaly protocols unseen during training (darkness, blur, actuation bias, processing latency) reduce average expected calibration error from 0.184 to 0.116, a 37% improvement over the best baseline, without modifying the base safety predictor.

Problem

• confidence calibration
• distribution shift
• anomaly detection
• vision-based safety prediction
• dynamics anomaly

Innovation

• confidence calibration
• anomaly detection
• vision-based safety prediction
• distribution shift
• world model