Towards Real-Time Graph Neural Network-Based 3D Object Detection for Autonomous Vehicles

3D object detection is an essential task for applications such as autonomous vehicles. Its ability to percept the surrounding environment in 3D space makes it more practical to implement in a real-world application. Light Detection and Ranging (LiDAR)-based 3D object detection gains its popularity and many approaches have been proposed to achieve state-of-the-art performance. On the other hand, Graph Neural Network (GNN)-based approach for 3D object detection is still in early development as compared to other approaches. Point-GNN is one of the first architectures to implement GNN in the Lidar-based 3D object detection task. While Point-GNN achieved leading accuracy in the KITTI detection benchmark, it has low inference speed which makes it not sufficient to be implemented in real-time applications such as an autonomous vehicle. In this work, we propose a method to improve the inference speed of Point-GNN by observing the contribution of graph scale and graph radius in the graph construction part of Point-GNN. The proposed approach improves the inference speed up to 2.83 times faster than the original Point-GNN while maintaining the degradation in overall Average Precision (AP) not exceeding 3.5%. In addition, we also observed the contribution of different features in the GNN layer to improve the overall AP by adding distance and local neighbor state as new additional features in the GNN layer. The proposed model gains up to 0.22% improved performance in BEV AP and 0.35%in the 3DAP.