Uncertainty Aware 3D Object Detection

Abstract

Object detection is one of the hottest topic right now and is a fundamental concept which determines to future of autonomous driving. There are hundreds of papers rolling out every day trying to improve the performances of these detectors by creating more complex models to increase the performance by a very little amount. The problem with current state-of-the-art approach is that the uncertainties of the model and the datasets are not being quantified. This is where the gap in the research lays as the uncertainties are not taken into account but also are not being used in the model itself. This research paper answers the question on how the uncertainties of a 3D LiDAR-based object detector can be quantified and used in 3D object detectors. This paper showcases that regarding semantic segmenta- tion and 3D object detection, that the multi-frame detectors have a great possibility of incorporating the uncertainties not only to improve the performances of the models but also to improve the quantification of the uncertainties. A key contribution of this paper is that it has shown that the best techniques for quantifying the uncertainties is to make use of a combination of MC dropout and variatonal inference to approximate the uncertainties, which can then be used to represent the spatial uncertainties in the detected objects. This research showcases that there is yet to be a paper which concisely uses these spatial uncertainties for multi-frame detectors. Furthermore, in order to create such a model, some of the most fundamental datasets have been analyzed (Kitti-360, Kitti, nuScenes, etc.) to show that the nuScenes, Waymo and Kitti-360 datasets are suitable for this kind of detection.