Road Segmentation Using UNet and ResNet-UNet Architectures: A Comparative Study

Authors:

Quentin Chappuis, Louis Martins and Kelu Huang

Dataset

The utilized dataset consists of 100 satellite images of roads along with their corresponding ground truth masks.

How to run?

1. Install PyTorch from here
2. Install the requirements as follows :

pip install -r requirements.txt

3. Extract the data from here and place them at the root of the repository. The training data should be in a folder named training and the test data in a folder named test_set_images
4. Run the run.py file
5. Get your predictions on the test set in the file submissions.csv placed at the root of the repository

Results

The model's performance on the test set achieved an F1 score of 0.829 and an accuracy of 0.908. This was accomplished using a ResNet-UNet architecture trained over 50 iterations with a learning rate of 1×10−4. For this project we used the Binary Cross-Entropy (BCE) loss.