Abstract
State-of-the-art 3D detection methods rely on supervised learning and large labelled datasets. However, annotating lidar data is resource-consuming, and depending only on supervised learning limits the applicability of trained models. Against this backdrop, here we propose using a self-supervised training strategy to learn a general point cloud backbone model for downstream 3D vision tasks. 3D scene flow can be estimated with self-supervised learning using cycle consistency, which removes labelled data requirements. Moreover, the perception of objects in the traffic scenarios heavily relies on making sense of the sparse data in the spatio-temporal context. Our main contribution leverages learned flow and motion representations and combines a self-supervised backbone with a 3D detection head focusing mainly on the relation between the scene flow and detection tasks. In this way, self-supervised scene flow training constructs point motion features in the backbone, which help distinguish objects based on their different motion patterns used with a 3D detection head. Experiments on KITTI and nuScenes benchmarks show that the proposed self-supervised pre-training increases 3D detection performance significantly.
Abstract (translated)
URL
https://arxiv.org/abs/2205.00705
