A Large-Scale Comprehensive Perception Dataset with High-Density Long-Range Point Clouds

News

Feb 22, 2021   Data released and open to download.

Features

All-Inclusiveness

We present a new autonomous driving dataset that includes all common sensors with all 360 degree coverage and has recorded diverse data with comprehensive annotation under various weather and lighting conditions. Our dataset is significantly different from existing driving datasets which usually miss more than one type of sensor or data variation or annotation that we have in a single dataset. We believe that our all-in-one driving dataset can be multi-purpose and benefit multi-task learning in autonomous driving, also enabling more opportunities of sensor fusion and higher robustness to rare data.

High-Density and Long-Range LiDAR

Our dataset is the first providing LiDAR data at a range of 1000 meters, which is about 10 times beyond the range of existing LiDAR sensors. In the meantime, we provide LiDAR data at a density of 1,000,000 points per frame, which is again about 10 times beyond the density of normal LiDAR data. We believe that our high-density and long-range LiDAR sensor enables more robust and early perception on small objects at far range, which is essential to motion planning and collision avoidance in autonomous driving.

Large-Scale Data

We collect the data at a large scale with an unprecedented number of the sequences each containing crowded agents, resulting in hundreds of millions of 2D and 3D annotations. To the best of our knowledge, our proposed dataset is the largest driving dataset so far, about 10 times larger than the Waymo open drive dataset. We believe that the large-scale and diverse data collected in our dataset can further improve the performance of perception and prediction models used for autonomous driving, especially when dealing with the rare driving situations.