Operational Readiness for Object Detection

Reliable object detection in safety-critical applications requires models that can recognize their own failures. This paper introduces a self-assessment approach for monitoring the operational readiness (OR) of object detectors under distributional shift. Unlike existing methods that focus on uncertainty estimation or generative models, we assess OR by comparing the cumulative distribution functions (CDFs) of feature map activations from multiple layers of a frozen detector. This captures low- and high-level features while providing a simple alternative to existing strategies. We formalize operational readiness as estimating the in-distribution probability of an input image as a proxy for data uncertainty. Our method is systematically compared to explicit generative modeling with normalizing flows and aggregation of detector uncertainties. Experiments on COCO with synthetic covariate shifts show that our activation-distribution approach outperforms these baselines. The method is detector-agnostic and can be added to pre-trained detectors.