Inria Aerial Image Labeling

Building Footprint Extraction using Deep Semantic Segmentation

To be able to use the code please follow listed instructions:

1. Fill in the form and download data from https://project.inria.fr/aerialimagelabeling/download/

2. Extract downloaded files and place them into data folder using the following folder structure:

   data/test/images/*.tif
   data/train/images/*.tif
   data/train/gt/*.tif
   

3. Execute prepare_data.py to image patches needed for training. The result would be the following folder structure:

   data/train_384x384/images/*.jpg
   data/train_384x384/gt/*.png
   

4. Execute train.py to initially train all 6 models. In case of an out of memory problem, adjust batch size in settings.py:

   batch_size = 9
   

5. Execute fine_tune.py to fine tune all 6 models. In case of an out of memory problem, adjust batch size in settings.py:

   batch_size = 9
   

6. Execute evaluate.py to evaluate fine-tuned models. The results will be placed in:

   tmp/eval_ft_1/*
   tmp/eval_ft_2/*
   tmp/eval_ft_3/*
   tmp/eval_ft_4/*
   tmp/eval_ft_5/*
   tmp/eval_ft_6/*
   

7. Execute prepare_submission.py to generate grayscale predictions for the test images. The results will be placed in:

   tmp/submission_grayscale/*
   

8. Execute grayscale_to_submission.py to prepare contest submission (requires GDAL). The results will be placed in:

   tmp/submission_0.45/*
   tmp/submission_0.45.zip