Spatially-Aware Reliability Modeling for BEV LiDAR 3D Vehicle Detection

Bird’s-eye-view (BEV) LiDAR detectors provide an efficient representation for real-time 3D perception, but their confidence scores may be imperfectly aligned with the localization quality of decoded 3D boxes. This score–localization mismatch can reduce the reliability of detection ranking, especially under strict IoU criteria and range-dependent LiDAR observations. To address this issue, we propose Spatially-Aware Quality Calibration (SAQC), a lightweight reliability-oriented scoring framework for BEV LiDAR 3D vehicle detection. SAQC estimates localization quality from coordinate-augmented local BEV feature patches around detected object centers, fuses the estimated quality with the raw detector score for quality-aware ranking, and applies post hoc sigmoid calibration with soft IoU targets for numerical quality alignment. Experiments on the KITTI Car validation split using an SFA3D-style center-based detector show that SAQC improves Moderate 3D AP from 89.13 to 89.57 at IoU = 0.7 and from 74.40 to 75.72 at IoU = 0.8. It also increases score–IoU Spearman correlation from 0.3685 to 0.4043 and reduces post-calibration Q-ECE from 0.0284 to 0.0183, while maintaining 103.6 FPS. These results indicate that local BEV spatial context can improve score–localization reliability without modifying decoded box geometry.