Sample grid how-to

sample_grid script performs a regular sampling over one or several zones.

This is one possibility for the first step (out of two) to build a dataset. This is suited if you have areas with ground truth for all your classes.

The goal is to generate a list of (x,y) coordinates :

• from each shape in a shapefile

• equally distributed (uniform distribution)

• allowing an (image,mask) to be completely inside the shape

Visualization of the result of a grid sampling :

The output contains as many files as there are shapes in the given shapefile. Each file is a list of coordinates from one shape. The shape are ordered in x, then y.

To launch the code, simply type odeon sample_grid -c <config.json>. You can add -v in order to have a verbose output.

Example :

$ odeon sample_grid -c ../tests/test_sampler_grid.json -v

Json file content

The full json file is :

{
  "image": {
    "image_size_pixel": 256,
    "resolution": 0.2
  },
  "sampler": {
    "input_file": "/home/data/33/ground_truth/2018/learning_zones/zone_all.shp",
    "output_pattern": "/media/hd/example_doc_sampling/normal/2018-*.csv",
    "shift": false,
    "tight_mode": false
  }
}

Image section

“image” describes the size and resolution of the future samples in the dataset:

• image_size_pixel (optional, integer, default value: 256): the number of pixels per side (usually a power of 2)

• resolution (optional, float or list of float, default value: 0.2): the resolution expressed in CRS unit, for instance 0.2 means 0.2m per pixel

This information has to be consistent with the following steps : generation of the dataset and training.

Sampler section

“sampler” contains information regarding the way the sampling will be performed:

• input_file (required, string): a file in any fiona accepted file format containing a list of shapes (avoid concave polygons)

• output_pattern (required, string or regex pattern): the directory where to store the output csv with a regex pattern option to set a prefix to the output files (useful to make distinction between different sampling output)

  Examples : “../../odeon_test/*_regularsampling.csv” and “/hd/odeon_test/regularsampling.csv” are valid inputs.

• strict_inclusion (optional, boolean, default value: false): use it if your extent or one of your extents is not squared, in this case it will only sample tiles strictly included in your extent.

• shift (optional, boolean, default value: false): this option will shift the offset of the grid representation (the upper left corner). It’s a useful option if you use the same extend in different moment of time, like in a multi-year training (combined with batch mode in the generate tool)

  • Example with shift :

    

  • Example without shift :

    

• tight_mode (optional, boolean, default value: false): tight every tile of the grid to the offset (upper left corner) when your extend is not a multiple of the desired tile size.

  • Example with tight_mode : there is no space between tiles and a big space on the right and bottom of the extent.

    

  • Example without tight_mode : the extent is not a multiple of the tile size and the rest is spread all over the grid.