Skip to content

deepbands/deep-learning-datasets-maker

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

64 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deep Learning Dataset Maker

We ❤️ Open Source

Python 3.8 QGIS 3.16.13 Code style: black license release

Deep Learning Datasets Maker is a QGIS plugin to make datasets creation easier for raster and vector data.

Run QGIS Desktop App (3.18) vi BinderHub! Click the button below to launch a server:

Binder

How to use

  1. Download and install QGIS and clone the repo :
git clone git@github.com:deepbands/deep-learning-datasets-maker.git
  1. Install requirements :

    • Enter the folder and install dependent libraries using OSGeo4W shell (Open As Administrator) :
    cd deep-learning-datasets-maker
    pip install -r requirements.txt
    • Or open OSGeo4W shell as administrator and enter :
    pip install Cython scikit-image Pillow pycocotools --user
  2. Copy folder named deep-learning-datasets-maker in QGIS configuration folder and choose the plugin from plugin manager in QGIS (If not appeared restart QGIS).

    • You can know this folder from QGIS Setting Menu at the top-left of QGIS UI Settings > User Profiles > Open Active Profile Folder .
    • Go to python/plugins then paste the deep-learning-datasets-maker folder.
    • Full path should be like : C:\Users\$USER\AppData\Roaming\QGIS\QGIS3\profiles\default\python\plugins\deep-learning-datasets-maker.
  3. Open QGIS, load your raster and vector data then select the output paths for rasterized, images and labels then click ok.

TODO

v0.2

  • Fix: If vector layer saved in memory not in file, rasterize can't work.
  • Splitting raster data into equal pieces with GDAL , https://gdal.org/.
  • Fix: Splitiing Image Size.
  • Rasterize shapefile to raster in the same satellite pixel size and projection.
  • Convert 24 or 16 bit raster to 8 bit.
  • Export as jpg (for raster) and png (for rasterized shapefile) with GDAL.
  • Converted semantic segmentation (0 and 1) to instance segmentation for labels (the original label is 0/255) option, and the result is a single-channel image that uses a palette to color.
  • PaddlePaddle Train/Val/Testing list text.
  • Use GDAL for instance segmentation instead of openCV.
  • Support COCO format.
  • Update plugin's UI :
    • Add new checkbox for other annotations like COCO.

v0.3

  • Fix : raster and vector full path on Linux/macOS (Sometimes cannot gdal/ogr.open from the full path because of forward slash /path_to_raster and backward slash \path_to_raster )