All-In-One Drive: A Comprehensive Perception Dataset with High-Density Long-Range Point Clouds

Xinshuo Weng, Yunze Man, Jinhyung Park, Ye Yuan, Matthew O'Toole, Kris M. Kitani

Keywords: Autonomous Driving, Perception, Datasets

Abstract: Developing datasets that cover comprehensive sensors, annotations, and out-of-distribution data is important for innovating robust multi-sensor multi-task perception systems in autonomous driving. Though many datasets have been released, they target different use-cases such as 3D segmentation (SemanticKITTI), radar data (nuScenes), large-scale training and evaluation (Waymo). As a result, we are still in need of a dataset that forms a union of various strengths of existing datasets. To address this challenge, we present the AIODrive dataset, a synthetic large-scale dataset that provides comprehensive sensors, annotations, and environmental variations. Specifically, we provide (1) eight sensor modalities (RGB, Stereo, Depth, LiDAR, SPAD-LiDAR, Radar, IMU, GPS), (2) annotations for all mainstream perception tasks (e.g., detection, tracking, prediction, segmentation, depth estimation, etc), and (3) out-of-distribution driving scenarios such as adverse weather and lighting, crowded scenes, high-speed driving, violation of traffic rules, and vehicle crash. In addition to comprehensive data, long-range perception is also important to perception systems as early detection of faraway objects can help prevent collision in high-speed driving scenarios. However, due to the sparsity and limited range of point cloud data in prior datasets, developing and evaluating long-range perception algorithms is not feasible. To address the issue, we provide high-density long-range point clouds for LiDAR and SPAD-LiDAR sensors (10x than Velodyne-64), to enable research in long-range perception. Our dataset is released and free to use for both research and commercial purpose: http://www.aiodrive.org/

URL: http://www.aiodrive.org/