Point Contrastive learning for LiDAR-based 3D object detection in autonomous driving

Efstathios Karypidis, Georgios Zamanakos, Lazaros T. Tsochatzidis, Ioannis Pratikakis

Abstract: Current progress in 3D Perception tasks for autonomous driving relies upon neural network architectures that their training requires a growing demand for annotated data. However, semantic annotation of 3D scenes is a very expensive and labor-intensive task. In this paper, we present an approach for self-supervised, data-efficient learning in the context of point contrastive learning, using two distinct pre-training techniques towards improving performance in LiDAR-based 3D object detection in autonomous driving. Our experimental work relies upon standard benchmarking datasets, namely KITTI and Waymo. Under a comprehensive evaluation framework it is shown that, in the absence of large annotated data, the proposed approach could achieve improved performance.