Road Extraction from Aerial Images

Delio Vicini, Matej Hamas, Taivo Pungas (Department of Computer Science, ETH Zurich, Switzerland)

The code in this repository trains a convolutional neural network and adds a post-processing layer, for the task of detecting roads in satellite images. See project report here.

Dependencies

• Python
• scipy
• numpy
• scikit-learn
• tensorflow
• matplotlib
• Pillow
• skimage

Setup

• Copy the data so the data/ folder contains the following:

test_set/
  |-- downsampled/
  |-- test_1.png
  |-- test_2.png
  |-- ...
  |-- test_50.png
  
training/
  |-- groundtruth/
  |   |-- downsampled/
  |   |-- satImage001.png
  |   |-- satImage002.png
  |   |-- ...
  |   |-- satImage100.png
  |
  |-- images/
      |-- downsampled/
      |-- satImage001.png
      |-- satImage002.png
      |-- ...
      |-- satImage100.png

• Run sh setup.sh while in the project root folder to set up the necessary file structures (assumes a Bash shell).

Running

Warning: running the code as-is requires around 100GB of memory.

• Run python run.py while in the src folder to train the CNN, apply post-processing and generate predictions.
• All results will be in the results folder, under CNN_Output/, postprocessing_output/ and in submission.csv.

Running the baseline

• Run python baseline/model_baseline.py while in the src folder.
• All results will be in the results folder, under CNN_Output_Baseline/ and in baseline_submission.csv.

The project has been tested with and is guaranteed to run on Python 3.5.0.