When Does Score Fusion Help? Conformally Certified Out-of-Distribution Detection for Camera and LiDAR Sensors

Area of Science:

• Computer Vision
• Sensor Fusion
• Machine Learning

Background:

• Autonomous systems rely on camera and LiDAR sensors, which are vulnerable to undetected distributional shifts.
• Existing out-of-distribution (OOD) detection methods lack comprehensive coverage of shift types and guarantees on false-positive rates (FPR).

Purpose of the Study:

• To investigate the effectiveness of calibrated score fusion for OOD detection in safety-critical autonomous systems.
• To develop a distribution-free finite-sample FPR certificate for OOD detectors.

Main Methods:

• Four post hoc scores (Maximum Softmax Probability, Energy, Mahalanobis distance, k-nearest-neighbour distance) were calibrated to p-values using the empirical cumulative distribution function (ECDF).
• Scores were combined using Fisher's method or cross-backbone z-score averaging and wrapped in a conformal predictor with Hoeffding-based Probably Approximately Correct (PAC) bounds.
• Evaluated on the PUG camera benchmark and nuScenes LiDAR dataset.

Main Results:

• Cross-backbone z-score averaging of Mahalanobis distances on camera data significantly improved AUROC (0.9258 to 0.9292).
• Uniform fusion provided limited gains on LiDAR data, with cross-backbone averaging showing substantial improvement on camera data.
• The distribution-free PAC certificate demonstrated low FPR margins (<1.5%) across both sensors.

Conclusions:

• Calibrated score fusion, particularly cross-backbone averaging, enhances OOD detection performance for camera sensors.
• The developed PAC certificate provides reliable FPR guarantees, crucial for safety-critical autonomous systems (e.g., ISO 26262, EASA CoDANN).
• Fusion strategies must consider sensor diversity for optimal out-of-distribution detection performance.