Inductive Conformal Prediction for Guaranteed Class-Label Coverage in Object Detection

Abstract

Conformal prediction has emerged as a principled framework for uncertainty quantification in computer vision, offering rigorous finite-sample coverage guarantees. However, its application in object detection has remained largely confined to localization, as standard inference codebases typically yield only top-1 class scores, precluding full class-label conformalization. In this work, we bridge this gap by adapting four architecturally diverse detectors—Faster R-CNN, RetinaNet, YOLO11, and RT-DETRv2—to facilitate the extraction of comprehensive per-class score vectors and the estimation of background confidence in the absence of native background modeling. Leveraging these adapted architectures, we implement inductive conformal prediction (ICP) using five distinct nonconformity functions: Top-K, Adaptive Prediction Sets (APS), Hinge, Margin, and Brier score. Our framework is rigorously benchmarked across a curated 20-class subset of MS-COCO and two specialized parasite egg datasets (AI4NTD P1.5v2 and Chula-ParasiteEgg-11). In addition, a Naive cumulative-threshold method is included as a baseline for comparison with APS, given their comparable mathematical formulations. Across target coverage levels of 90%, 95%, and 99%, the conformalized models consistently achieved nominal coverage with only minor finite-sample deviations. Hinge and APS exhibited an optimal balance between statistical coverage and prediction-set efficiency, whereas Margin and Brier scores tended toward larger sets under high data complexity and strict coverage requirements. With empty prediction sets maintained below 0.1%, our findings establish ICP as a robust and adaptable paradigm for trustworthy class-label uncertainty estimation, particularly within safety-critical workflows such as automated parasite diagnostics.