LGNet: Local and global point dependency network for 3D object detection

Abstract

3D object detection is a challenging task in autonomous driving industry scenarios. Many pre-existing methods employ the set-abstraction operation for generating key-point representations, which, however, cannot learn the long-range context dependency properly. In addition, the pooling operator, which only focuses on maximum channel response, is adopted to aggregate features of neighbor points without semantic information. To fix these issues, we propose LGNet, a new framework that simultaneously captures local and global point dependencies for enhancing 3D object detection. Specifically, we first introduce a new local point-graph pooling module to compute point-to-point correlations in a local region and aggregate features from neighboring points. To further capture the long-range dependency in a global context, we devised a global point-aware module to integrate local and global features at higher resolution. Experiments on the KITTI 3D detection dataset and Waymo Open Dataset benchmark show that LGNet achieves state-of-the-art performance in multiple classes. We will upload the code on https://github.com/MWPony/LGNet.