Motional releases expanded autonomous driving dataset with over 1.4 billion annotated lidar points

A year ago, Scale and NuTonomy released the autonomous driving dataset NuScenes, claiming that the dataset surpassed datasets such as KITTI, Baidu ApolloScape, and Udacity autonomous vehicle data in both scale and accuracy. Since then, a variety of new and more diverse datasets have emerged in the industry, such as Waymo Open Dataset, Ford's autonomous vehicle dataset, and Lyft's autonomous vehicle dataset. According to foreign media reports, Motional (whose CEO founded NuTonomy) recently released an expanded version of the NuScenes dataset.

(Image source: Motional)

Datasets like NuScenes can be used to improve the robustness of self-driving cars in a variety of environments. The Rand Corporation, a U.S. research organization, estimates that self-driving cars need to drive 11 billion miles to obtain reliable safety data, but because some difficulties slow down the pace of testing in the real world, simulated driving miles become the best choice.

The expansion of NuScenes includes NuScenes-lidarseg, which improves the semantic segmentation of 1,000 Singapore and Boston scenes, making it one of the largest open lidar segmentation datasets. According to Motional, NuScenes-lidarseg adds 1.4 billion annotated lidar points, providing a more detailed picture of the vehicle environment than the original bounding box, allowing researchers to study lidar point cloud segmentation and foreground extraction.

The expanded dataset also includes a new dataset, NuImages, which contains nearly 100,000 annotated 2D images to represent a range of challenging and "educational" driving conditions. Motional said that NuImages was created based on user needs to help autonomous vehicles drive safely in unpredictable scenarios.

Both NuScene-lidarseg and NuImages are based on the existing NuScenes dataset, which contains hundreds of scenes, including more than 1 million images captured by cameras, lidar, radar, GPS, and inertial measurement sensors. Motional said that since its release in March 2019, more than 8,000 researchers have used NuScenes, more than 10 new datasets have been made public, and more than 250 scientific papers have cited its data.