Detecting Roads in Aerial Images using Deep Neural Networks

We implement a classifier following a U-Net architecture that outperforms more basic approaches and segments unseen data with high accuracy. Our model was trained using the Pytorch framework on Google's Colab infrastructure.

Contents

• The src folder contains all our models along with helpers for metric and loss functions.
• The notebooks used to train CNN and UNet.
• A run script to reproduce the results of our best AICrowd submission (more on that below).
• A pre-trained model pretrained_model.pkt.

Setup

The following is a list of dependencies used in this study. You can install them using either pip or anaconda:

• Pytorch, version 1.7.0
• Pytorch learning rate finder, version 0.2.1
• Matplotlib, version 3.2.2
• Numpy, version 1.18.5
• Scikit-image, version 0.17.2

To run the various notebooks, you will need jupyter lab (or notebook). Note that we have moved the files during development so the imports migth be broken.

Training instructions

In the root folder, you will find the run.py script. A the top of the file, there are 3 varriables :

• load_pretrained_model, a boolean that specifies whether to load a pretrained model or train a new one from scratch (default is True).
• pretrained_model_path, the path to the pretrained model. If the the previous varriable is set to False then the resulting trained model will be saved at that path.
• root_data_path, the path to the directory containing all the training and testing data.

When the program is done training (or loading the model), it writes, in a outputs directory, the prediction masks of our model on the testing images. Using this directory, it creates the submission file submission.csv.

Authors

• Karim Assi ([email protected])
• Alexandre Pinazza ([email protected])
• Milan Reljin ([email protected])