Uncertainty Estimation for Monocular 3D Object Detectors in Autonomous Driving
Qinghua Yang, Hui Chen, Zhe Chen, Junzhe Su · 2021
Uncertainty estimation for 3D object detectors plays a critical role in autonomous driving. This is because current state-of-the-art 3D object detectors can make severe errors in their detections, and not knowing the uncertainties in these results may cause catastrophic consequences for safety-critical autonomous vehicles. Prior researches have studied the uncertainty estimation problem for 2D and lidar-based 3D object detectors, but little attention has been paid to monocular 3D object detectors. In consideration of the extensive need to achieve 3D object detection with low-cost cameras, it is of great importance to perform uncertainty estimation for monocular 3D object detectors as well. Thus, in this paper, we propose a merging strategy to perform sampling-based uncertainty estimation for monocular 3D object detectors. Specifically, we adopt the popular Monte Carlo Dropout method to obtain the sampling results for uncertainty estimation, and propose to merge them by soft clustering and Bayesian Inference in the Bird's Eye View. During this process, in order to reduce computational cost of multiple sampling times, we propose the weighted Monte Carlo Dropout uncertainty calculation method that is able to capture uncertainties from very few sampling results. Finally, we verify the effectiveness of the proposed method and its superiority to the existing method on the challenging KITTI dataset.