OpenScene is the largest scale autonomous driving 3D occupancy prediction benchmark in the world, containing 120 hours and 4 million frames of multi-sensor data, covering urban scenes in Las Vegas, Singapore, Boston, Pittsburgh, and more, suitable for autonomous driving perception and planning research.
The world's largest 3D occupancy prediction benchmark, providing a standardized evaluation platform for autonomous driving perception research.
Integrates LiDAR point clouds and multi-view camera images, providing complete 3D perception input to support research and development of multi-modal fusion algorithms.
Provides detailed 3D voxel occupancy labels, discretizing the scene into a regular grid and annotating the semantic category and occupancy state of each voxel.
Data collected from Las Vegas, Singapore, Boston, and Pittsburgh, covering diverse road structures and traffic scenarios.
Includes 120 hours of continuous driving data, covering different times, weather, and lighting conditions, providing rich training material for temporal modeling.
Provides multi-level annotations such as 3D bounding boxes, semantic segmentation, and occupancy grids, supporting various downstream tasks like detection, segmentation, and prediction.
Provides unified evaluation metrics and leaderboards, facilitating fair comparison of the performance of different 3D occupancy prediction methods for researchers.
From academic research to industrial implementation, supporting the full chain of perception development for autonomous driving
Train and evaluate voxel-level occupancy prediction models, understanding the complete geometry and semantic structure of the three-dimensional space around the vehicle
Provide training data for autonomous driving perception systems, supporting end-to-end perception, planning, and decision-making algorithm development
Perform 3D semantic segmentation and instance segmentation from multi-sensor data, deeply understanding complex traffic scenes
Research LiDAR-Camera cross-modal fusion strategies to enhance 3D perception accuracy and robustness
The following is an example of typical sensor configuration and annotation format for the OpenScene dataset
{
"scene_token": "a1b2c3d4e5f6...",
"city": "las_vegas",
"duration_s": 20.0,
"num_frames": 40,
"sensors": {
"lidar": {
"type": "lidar",
"channel": "LIDAR_TOP",
"num_points_per_frame": 300000
},
"cameras": [
{"channel": "CAM_FRONT", "resolution": [1600, 900]},
{"channel": "CAM_FRONT_LEFT", "resolution": [1600, 900]},
{"channel": "CAM_FRONT_RIGHT", "resolution": [1600, 900]},
{"channel": "CAM_BACK", "resolution": [1600, 900]},
{"channel": "CAM_BACK_LEFT", "resolution": [1600, 900]},
{"channel": "CAM_BACK_RIGHT", "resolution": [1600, 900]}
]
},
"occupancy_grid": {
"voxel_size": [0.5, 0.5, 0.5],
"grid_range": [[-50, -50, -5], [50, 50, 3]],
"num_classes": 16
}
}
From browsing to loading, you can start your 3D perception research in just a few minutes
View dataset details on the Ace Data Cloud platform to understand metadata such as sensor configuration, annotation format, and scene distribution.
Select the data slices for the corresponding city and scene based on research needs, supporting on-demand downloads of LiDAR point clouds, camera images, and occupancy annotations.
Use the nuScenes devkit to load the data for 3D visualization, model training, and occupancy prediction evaluation.
The world's largest autonomous driving 3D occupancy prediction benchmark, open license, available for immediate download. Whether you are a perception algorithm researcher or an autonomous driving engineer, this dataset is an ideal choice.