Rethinking Detection Calibration: A Coordinate and Direction Perspective

2026-07-31

AI Summary

This work addresses the limitation of existing object detection calibration methods, which rely solely on overall bounding box localization accuracy and thus fail to accurately characterize the confidence of individual coordinate predictions. To overcome this, the paper proposes ReDC, a coordinate-level calibration framework that introduces coordinate alignment and directional bias modeling for the first time. By recoding coordinate-wise confidence, estimating offset directions, and aggregating these into a box-level representation, ReDC enables fine-grained assessment of localization confidence. The method remains compatible with conventional calibration formulations and consistently outperforms current approaches in both in-domain and out-of-domain experiments, significantly improving the accuracy of localization uncertainty estimation.

Abstract

Deep learning based object detectors require trustworthiness beyond competitive detection performance, but deep neural networks are prone to overconfident predictions, assigning high confidence scores to predictions that are likely to be inaccurate. To improve the alignment between confidence scores and prediction accuracy, existing methods calibrate confidence scores based on box-level localization, such as precision or intersection over union with the ground truth bounding box. However, box-level localization reflects only a measure of agreement between the predicted box and the ground truth, resulting in calibrated confidence scores for box-level accuracy failing to capture the localization accuracy of coordinates of box. To tackle this issue, we propose a novel post-hoc calibration framework, rethinking detection calibration (ReDC), which provides reliable coordinate-level confidence scores, including directional information. The proposed framework defines coordinate-wise alignment and deviation direction between predictions and ground truth. Based on the alignment measure, confidence re-encoding produces reliable coordinate-level confidence scores, while directional displacement estimation predicts coordinate-wise deviation directions. Extensive experiments under in-domain and out-domain scenarios demonstrate that the proposed approach expresses the coordinate-wise localization of detected objects more precisely than existing methods. Furthermore, our method covers the representational scope of prior calibration approaches by aggregating coordinate-level confidence scores into box-level localization.

Problem

• detection calibration
• localization accuracy
• confidence score
• coordinate-level
• object detection

Innovation

• detection calibration
• coordinate-level confidence
• directional displacement
• post-hoc calibration
• localization accuracy