Can We Trust You? On Calibration of a Probabilistic Object Detector for Autonomous Driving
It addresses a critical safety issue in autonomous driving perception systems by improving uncertainty calibration, though it is incremental as it builds on existing detection methods.
The paper tackled the problem of miscalibrated uncertainty estimates in probabilistic LiDAR 3D object detectors for autonomous driving, proposing three methods that significantly reduced calibration errors and generalized well across datasets.
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. Extensive experiments on several datasets show that our methods produce well-calibrated uncertainties, and generalize well between different datasets.