Skip to content

Latest commit

 

History

History
195 lines (168 loc) · 9.41 KB

examples.md

File metadata and controls

195 lines (168 loc) · 9.41 KB
Raw

Examples

Example Data:

All examples provided under <examples> assume that the input data has a similar structure to

├── <example_dataset>
│   ├── raster
│   │   ├── image_1
│   │   │   ├── image_1.tif
│   │   ├── image_1-labels
│   │   │   ├── image_1.tif
│   │   ├── image_2
│   │   │   ├── image_2.tif
│   │   ├── image_2-labels
│   │   │   ├── timage_2.tif
│   │   ├── ...
├── ...

The examples use either the ISPRS Potsdam or the Open Cities AI dataset. The instructions to prepare both datasets are located here and here. The Potsdam dataset has smaller file sizes and more category labels and is therefore more suitable for testing. Only the provided rasterization examples require the geojson files contained in the Open Cities AI dataset.

By running one of the various examples located in <examples> the pipeline will create additional (intermediate) results in this directory, e.g.

├── <examples_dataset>
│   ├── raster
│   │   ├── image_1
│   │   │   ├── image_1.tif
│   │   ├── image_1-labels
│   │   │   ├── image_1.tif
│   │   ├── image_2
│   │   │   ├── image_2.tif
│   │   ├── image_2-labels
│   │   │   ├── timage_2.tif
│   ├── image_tiles
│   ├── label_tiles
│   ├── predicted_tiles
│   ├── fused_tiles
│   ├── fused_tiles_aggregated
│   ├── comparison_tiles
│   ├── ...
├── ...

Earth Observation Tile (EOT) examples using the ISPRS Potsdam dataset:

The following examples demonstrate the core capabilities of the EOT library using the raster data of the ISPRS Potsdam dataset.

EOT Color Tiling (eot_color_tiling_geojson_example.py)

  • Divides all RGB raster images into EOT rgb tiles
  • Creates a directory <examples_dataset>/image_tiles containing the EOT rgb tiles
  • Creates for each input raster image a corresponding grid.json file representing the created tiles as geojson.

EOT Color Tiling with tile specific geo-data (eot_color_tiling_aux_xml_example.py)

  • Divides all RGB raster images into EOT rgb tiles (similar to eot_color_tiling_geojson_example.py)
  • Creates a directory <examples_dataset>/image_tiles containing the EOT rgb tiles
  • Creates for each tile an aux.xml file with the corresponding geo-data. When opening the jpg tile in QGIS, QGIS picks up the geo-information from the corresponding aux.xml file to correctly visualize the tile information

EOT Label Tiling (eot_label_tiling_geojson_example.py)

  • Divides all label raster images into EOT label tiles
  • Creates a directory <examples_dataset>/label_tiles containing the label rgb tiles (uses a color palette to represent the classes)
  • Creates for each input raster image a corresponding grid.json file representing the created tiles as geojson.
  • Creates for each input raster image and each dataset category a file with the aggregated category information (such as building.json, car.json, ...)

EOT Color Fusion (eot_color_fusion_example.py)

  • Performs the full pipeline (similar to eot_label_fusion_example.py), but uses the RGB raster images to visualize the tiling and the fusion step.
  • Creates the following directories:
    • <examples_dataset>/image_tiles
    • <examples_dataset>/predicted_tiles
    • <examples_dataset>/fused_tiles
    • <examples_dataset>/fused_tiles_aggregated
    • <examples_dataset>/comparison_tiles
    • <examples_dataset>/comparison_tiles_aggregated
  • The tile prediction step performs a dummy segmentation which draws a rectangle on each rgb tile denoting the corresponding reliable area.
  • The fused tiles in <examples_dataset>/fused_tiles show the fused results for each tile containing the reliable areas of overlapping tiles.
  • The aggregated fused tiles in <examples_dataset>/fused_tiles_aggregated show the composition of the reliable areas of all tiles relative to the tiling of the raster image.

EOT Label Fusion (eot_label_fusion_example.py)

  • Performs the full pipeline (tiling, tile fusion, tile aggregation) using the label raster images
  • Creates the following directories:
    • <examples_dataset>/label_tiles
    • <examples_dataset>/predicted_tiles
    • <examples_dataset>/fused_tiles
    • <examples_dataset>/fused_tiles_aggregated
    • <examples_dataset>/comparison_tiles
    • <examples_dataset>/comparison_tiles_aggregated
  • The tile prediction step performs a dummy segmentation, i.e. it draws a rectangle on each tile denoting the corresponding reliable area (using a new index in the color palette)
  • The fused tiles in <examples_dataset>/fused_tiles show the fused results for each tile containing the reliable areas of overlapping tiles.
  • The aggregated fused tiles in <examples_dataset>/fused_tiles_aggregated show the composition of the reliable areas of all tiles relative to the tiling of the raster image.

Web Map examples using the ISPRS Potsdam dataset:

In addition to EOT tiles the library also supports web tiles (based on a mercator projection) as used by services such as OpenStreetMap. The following examples also use the raster data of the ISPRS Potsdam dataset. Web map tiles do not allow to fuse semantic segmentations of adjacent tiles.

Web Color Tiling (web_color_tiling_geojson_example.py)

  • Creates a directory <examples_dataset>/image_tiles containing the web rgb tiles (similar to eot_color_tiling_geojson_example.py)
  • The web tiles base on a spherical mercator projection of the data
  • Creates a single grid.json file representing the created tiles as geojson (this differs from the EOT approach where a grid is created for each input raster image)

Web Label Aggregation (web_label_tiling_aggregation_example.py)

  • Creates a directory <examples_dataset>/label_tiles containing the web label tiles (similar to eot_color_tiling_geojson_example.py)
  • The web tiles base on a spherical mercator projection of the data
  • Creates <examples_dataset>/tiles_aggregated containing the aggregated results of the web label tiles using geojson files and raster images

Rasterization examples using the Open Cities AI dataset:

The pipeline usually uses raster images to represent the label data, since this offers several advantages such as a tight coupling of category information and geo-data matching the properties of the corresponding color images. In order to convert vector labels (e.g. geojson files) to raster labels see this example. Converting vector labels to raster images before tiling avoids the repetition of rasterization step while testing different tiling schemes. In case you want to work with the vector labels directly, see the following rasterization examples that use the geojson files of the Open Cities AI dataset.

EOT Rasterization Aggregation (eot_label_rasterization_aggregation_example.py)

  • Create <examples_dataset>/label_tiles_rasterized with EOT label tiles derived from a geojson label files (instead of label raster images).
  • Creates <examples_dataset>/rasterized_tiles_aggregated with the corresponding aggregated results

EOT Tiling Aggregation (eot_label_tiling_aggregation_example.py)

  • Produces a similar result to eot_label_rasterization_aggregation_example.py but uses raster labels in the tif images (instead of the vector labels in the geojson files)

Web Rasterization Aggregation (web_label_rasterization_aggregation_example.py)

  • Create <examples_dataset>/label_tiles_rasterized with web label tiles derived from a geojson label files (instead of label raster images).
  • Creates <examples_dataset>/rasterized_tiles_aggregated with the corresponding aggregated results