その不確かさを信頼できるか?自動運転向け確率的物体検出器のキャリブレーション
Can We Trust You? On Calibration of a Probabilistic Object Detector for Autonomous Driving
LiDAR 3D物体検出ネットワークにおける不確かさのキャリブレーション誤差を指摘し、それを大幅に低減する3つの実用的手法を提案した論文。
著者: Di Feng, Lars Rosenbaum, Claudius Glaeser, Fabian Timm, Klaus Dietmayer
分類: cs.RO, cs.CV
原文アブストラクト
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.
