Self-Supervised Pretraining for Point Cloud Object Detection in Autonomous Driving

Abstract

Autonomous driving promises a safer and more efficient means of transportation. However, one of many challenges it faces is understanding the complex driving environment. Modern perception systems often utilize neural networks and tackle complicated driving scenarios in a data-driven manner. Unfortunately, such systems require a large amount of labeled data, which can be prohibitively expensive to collect. In this paper, we investigate self-supervised learning as a method to reduce the reliance on labeled data in the context of autonomous driving. We specifically focus on point cloud recognition and apply contrastive and geometric pretext tasks to pretrain neural networks using unlabeled point cloud data. We conduct experiments in the nuScenes autonomous driving dataset with various amounts of labeled data. Our experiments reveal three insights: (1) pretraining with contrastive loss alone improves the average precision (AP) but negatively impacts the object heading accuracy, (2) combining contrastive and geometric pretext tasks benefits both the average precision and heading accuracy, and (3) the improvement by self-supervised pretraining remains even with an increased amount of labeled data and training steps.