README.md

Storm drains dataset

pointclouds

This dataset contains 373 HD 3d lidar pointclouds. Each pointcloud is a 2mx2m tile.

Among those pointclouds, 182 contains at least one storm drain and 191 none.

Those pointclouds are split into a training set (223 pointclouds) and a testing set (150 pointclouds).

features and labels

Each point has the following properties:

• x, y, z 3d coordinates in EPSG:2950

• intensity of the signal from the lidar sensor

• label which has the value 1 if that point belongs to a storm drain (0 otherwise).

• some computed geometric features given its neighborhood: The neighborhood is computed as a sphere of ray r which can be 3cm, 5cm, 10cm or 20cm. The geometric feature will be prefixed with __0_03 for the sphere of 3cm, __0_05 for the sphere of 5cm, etc...

  • nb_neighbors
  • pca1
  • pca2
  • anisotropy
  • eigen_entropy
  • eigenvalue_sum
  • linearity
  • nx
  • ny
  • nz
  • omnivariance
  • planarity
  • sphericity
  • surface_variation
  • verticality

  Those geometric features are described in the following references :

  • Brodu, N., & Lague, D. (2012). 3D terrestrial lidar data classification of complex natural scenes using a multi-scale dimensionality criterion: Applications in geomorphology. ISPRS Journal of Photogrammetry and Remote Sensing, 68, 121-134. pdf
  • Weinmann, M., Jutzi, B., & Mallet, C. (2017). Geometric features and their relevance for 3d point cloud classification. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 4, 157. pdf

from jakarto_datasets.datasets.storm_drains import StormDrainsDataset

# Load dataset
storm_drains_2019 = StormDrainsDataset()

for data in storm_drains_2019.training_set:
    coordinates = data.get_coordinates_data()  # x, y, z in EPSG:2950
    lidar_data = data.get_lidar_data()  # intensity, gps_time
    extra_lidar_data = data.get_extra_lidar_data()  # geometric features

    label = data.get_label_lidar_data()  # labels

    print(data)
    print(coordinates.shape)
    print(lidar_data.shape)
    print(extra_lidar_data.shape)
    print(label.shape)

rasters and masks

We also rasterized the lidar pointclouds to benefit of Computer Vision algorithms.

from jakarto_datasets.datasets.storm_drains import StormDrainsDataset

# Load dataset
storm_drains_2019 = StormDrainsDataset()

for data in storm_drains_2019.training_set:
    raster = data.get_raster()
    mask = data.get_mask()

    print(data)
    print(raster.shape)
    print(mask.shape)

Examples

Please refer to the following examples :

• A simple viewer to loop over the dataset. It allows to visualize the pointcloud, the raster and the mask at the same time.
• train a machine learning model given only intensity and computed geometric features, try to predict the label for each point of pointclouds.