DCGNN: a single-stage 3D object detection network based on density clustering and graph neural network

Abstract Currently, single-stage point-based 3D object detection network remains underexplored. Many approaches worked on point cloud space without optimization and failed to capture the relationships among neighboring sets. In this paper, we propose DCGNN, a novel based density clustering graph neural networks. DCGNN utilizes ball query partition exploits local global by Density optimizes partitioned original approach ensure key sets containing more detailed features of objects. Graph networks are very suitable for exploiting points Additionally, as network, achieved fast inference speed. We evaluate our KITTI dataset. Compared with state-of-the-art approaches, proposed better balance between performance time.