Exploring local delaunay graph based neural network for 3D object detection

Abstract

Three-dimensional object detectors are essential components of the perception sys tems of autonomous robots. Most of the 3D object detection methods are point-based, voxel-based, or point-voxel-based (uses both point-based and voxel-based approaches). Recently, few works have used graph neural networks for 3D object detection and achieved promising results. Representation of a point cloud as a graph preserves the irregularity and removes the cost associated with the transformation into voxel grids or projection into a bird’s eye view. In this thesis, we study the effects of point cloud encoding using local Delaunay graphs by embedding it in an existing graph neural network (Point-GNN). We perform experiments on the KITTI benchmark and find that graph-based methods achieve higher or comparable accuracy. Our results are comparable with most of the state-of-the-art methods. Further, our study indicates that the selection of the right type of proximity graph representation is crucial for real-time 3D object detection.