Can We Trust You? On Calibration of a Probabilistic Object Detector for Autonomous Driving

Understand

Reliable uncertainty estimation is crucial for perception systems in safe autonomous driving.

• Recently, many methods have been proposed to model uncertainties in deep learning based object detectors.
• However, the estimated probabilities are often uncalibrated, which may lead to severe problems in safety critical scenarios.
• In this work, we identify such uncertainty miscalibration problems in a probabilistic LiDAR 3D object detection network, and propose three practical methods to significantly reduce errors in uncertainty calibration.