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

TL;DR

This paper addresses the calibration issues in probabilistic LiDAR 3D object detectors for autonomous driving, proposing methods to improve uncertainty estimation accuracy for safer perception systems.

Contribution

It identifies calibration problems in existing probabilistic detectors and introduces three practical methods to enhance uncertainty calibration and generalization.

Findings

Proposed methods significantly reduce calibration errors.

Enhanced calibration improves safety in autonomous driving scenarios.

Methods generalize well across multiple datasets.

Abstract

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.