Datasets
A quick summary of the major 3D-object-detection datasets used as benchmarks in autonomous-driving research. Each dataset was collected by equipping a vehicle with multiple cameras and LiDAR sensors, driving through varied environments, and then manually annotating every 3D frame.
| Dataset | Year | # Cameras | # LiDARs | # Scenes | # Classes | Locations | Night | Rain | Annotated 3D BBoxes | Annotated Frames |
|---|---|---|---|---|---|---|---|---|---|---|
| KITTI [Geiger et al. 2013] | 2012 | 2 | 1 | 22 | 3 | Germany | No | No | 80 000 | 15 000 |
| ApolloScape [Wang et al. 2019] | 2018 | 2 | 2 | 73 | 27 | China | Yes | No | 70 000 | 80 000 |
| nuScenes [Caesar et al. 2020] | 2019 | 6 | 1 | 1 000 | 23 | USA / Singapore | Yes | Yes | 1.4 M | 40 000 |
| Argoverse [Chang et al. 2019] | 2019 | 9 | 2 | 113 | 15 | USA | Yes | Yes | 993 000 | 22 000 |
| Waymo Open [Sun et al. 2019] | 2019 | 5 | 5 | 1 150 | 4 | USA | Yes | Yes | 12 000 000 | 230 000 |
| Lyft Level 5 [Houston et al. 2021] | 2019 | 7 | 3 | 366 | 9 | USA | No | No | 1.3 M | 46 000 |
| H3D [Patil et al. 2019] | 2019 | 3 | 1 | 160 | 8 | USA | No | No | 1.1 M | 27 000 |
Parameter Descriptions
- Dataset: Name of the benchmark and original publication citation.
- Year: Release year of the dataset.
- # Cameras: Number of synchronized RGB cameras.
- # LiDARs: Number of LiDAR sensors.
- # Scenes: Distinct driving sequences or collection drives.
- # Classes: Number of object categories annotated (e.g., cars, pedestrians).
- Locations: Geographic areas where data was recorded.
- Night: Indicates if night-time data is included.
- Rain: Indicates if driving in rainy conditions is present.
- Annotated 3D BBoxes: Total count of 3D bounding boxes labeled.
- Annotated Frames: Number of keyframes with full 3D annotations.