Benchmark Analysis of Deep Learning-based 3D Object Detectors on NVIDIA Jetson Platforms

Abstract

3D object detection could be highly beneficial for autonomous driving of mobility platforms such as robots and drones. Since the function provides a one-shot inference that extracts 3D position with depth information and heading direction of neighboring objects, robots can generate a reliable path to navigate without a collision. In order to enable the smooth functioning of 3D object detection, there have been several approaches to build detectors using deep learning for fast and accurate inference. In this paper, we investigate 3D object detection frameworks and analyze their performance on Jetson boards released by NVIDIA. Since mobility platforms often require real-time control to avoid dynamic obstacles, onboard processing with a built-in computer is an emerging trend. Jetson series solve such requirement with a lightweight size of the board and a suitable computational performance for autonomous navigation. Recently, many robotic applications using the Jetson series have been studied owing to their clear benefits. To examine the Jetson series for a computationally expensive task like point cloud processing, we test the performance of the boards (i.e., AGX and nano) using deep learning-based 3D object detectors. We present benchmark results in terms of three metrics including accuracy, frame per second (FPS), and resource usages.