PV-GNN:Point-Voxel 3D Object Detection based on Graph Neural Network

Abstract Recently, two-stage detectors have been widely used in 3D object detection.Most of them perform ROI feature extraction using keypoint sampling or re-voxelization in the second stage.However, these methods do not make full use of the point correlation and ignore the surface texture information hidden in the point cloud.In this paper, we propose a novel 3D object detection framework, PointVoxel-GNN (PV-GNN), to accurately detect 3D objects from point cloud data.The framework uses both voxel neural network and PointNet-based set abstraction to extract more diverse point cloud features, and uses graph neural network to fully exploit the correlation of point cloud data.Specifically, the framework firstly converts the sparse point cloud data into a regular voxel network, and extracts effective features through the neural network. Then, all the voxels of the whole scene feature volume are summarized into a small number of feature keypoints by using the voxel-to-keypoint scene coding module.In order to further generate high-quality proposals, a GNN-RoI pooling layer is proposed to explicitly extract the texture information of the object surface through the graph neural network.Compared with the traditional pooling operation, GNN-RoI can encode richer context information, which is used to accurately estimate the confidence and location of the object.The framework has conducted a large number of experiments on KITTI dataset and Waymo Open dataset, and our proposed PV-GNN redetection accuracy has achieved significant progress.