Our Baraja 3D Car Detection dataset uses the Baraja Spectrum-Scan lidar, which is a high resolution, adjustable scan pattern lidar. The design of the lidar allows it to adjust its vertical and horizontal angular resolution for user-defined non-uniformly spaced subdivisions. This dataset uses 128 non-uniform subdivisions per channel of the Baraja Spectrum-Scan which is a higher density of point cloud compared to many public point cloud datasets such as KITTI or Waymo Open Dataset. The scan pattern is also unique due to the non-uniformity of the vertical subdivisions. There are two channels for the Baraja lidar which have a slight vertical offset that allows it to achieve an interleaved pattern to give more useful 3D information. We provide calibrated and time-synced frames with an Intel RealSense D435i camera. The dataset was collected in San Francisco in dry weather conditions and contains 100 frames with a total of 928 cars.


When using this dataset please cite:

Tsai, D., Berrio, J. S., Shan, M., Worrall, S., & Nebot, E. (2021). See Eye to Eye: A Lidar-Agnostic 3D Detection Framework for Unsupervised Multi-Target Domain Adaptation. arXiv preprint arXiv:2111.09450.

Download from Local site