If you use one these data sets, or results published here or in the original paper, please cite:

Dirk Holz and Sven Behnke. Registration of Non-Uniform Density 3D Point Clouds using Approximate Surface Reconstruction, in Proceedings of the International Symposium on Robotics (ISR) and the German Conference on Robotics (ROBOTIK), München, Germany, 2014.

  @InProceedings{holz_mav_mesh_reg,
    author =       {Dirk Holz and Sven Behnke},
    title =        {Registration of Non-Uniform Density 3D Point Clouds
                    using Approximate Surface Reconstruction},
    booktitle =    {Proceedings of the International Symposium on
                    Robotics (ISR) and the German Conference on Robotics 
                    (ROBOTIK)},
    year =         {2014},
  }
        

The data was recorded using the a continuously spinning laser scanner on a mobile ground robot standing still while acquiring 3D point clouds (thus avoiding inaccuracies in laser scan aggregation). The dataset contains point clouds from eight different poses with a total of 6890 2D laser scans acquired over multiple full rotations at each pose to obtain comparably dense point clouds. The total trajectory length between the eight poses is roughly 50m. It was recorded by Schadler et al. [1] in the arena of the DLR SpaceBot Cup competition for semi-autonomous exploration and mobile manipulation in rough terrain. For the data set, we collected all 2D scan lines acquired at each of the poses, sorted them by rotation angle and re-organized the data to obtain eight full resolution (θ ≈ 0.3°) organized point clouds. We annotated each point cloud with the ground truth pose estimate obtained from an accurate multi-resolution surfel mapping approach for dense point clouds [1]. For the experimental evaluation, we generated thinned out versions of these eight original point clouds with different angular resolutions and angles θ ∈ [1°, 90°], respectively.

References

1. Mark Schadler, Jörg Stückler, and Sven Behnke.
   Rough Terrain 3D Mapping and Navigation using a Continuously Rotating 2D Laser Scanner.
   Accepted for German Journal on Artificial Intelligence (KI), Springer, to appear 2014.
2. Dirk Holz and Sven Behnke.
   Fast range image segmentation and smoothing using approximate surface reconstruction and region growing.
   In Proceedings of the International Conference on Intelligent Autonomous Systems (IAS), 2012.

The data set was recorded with the continuously spinning laser scanner on the flying MAV. The MAV was flying through a parking garage of 40m x 20m. Overall, the dataset contains a total of 4420 2D scan lines which are aggregated to 200 3D scans (each aggregated over one half rotation of the scanner). The overall trajectory length is 73m (traveled in 100s). We used two fish-eye stereo camera pairs on the MAV and visual odometry [1] to estimate the MAV's motion and aggregate the individual 2D laser scans to 3D point clouds.

References

1. Johannes Schneider, Thomas Läbe and Wolfgang Förstner.
   Incremental real-time bundle adjustment for multi-camera systems with points at infinity.
   In ISPRS Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, volume XL-1/W2, 2013.