Machine Learning: Project 2 - Road Segmentation

Project 2 of the Machine Learning course given at the EPFL Fall 2021.

The goal of this project is to segment satellite images by detecting roads. Our classifier consists of a convolutional neural network called UNet.

Team members

• Quentin Deschamps
• Emilien Seiler
• Louis Le Guillouzic

Results

Model	Data augmentation	Postprocessing	F1 score	Accuracy	Submission
UNet	Yes	Yes	0.901	0.946	#169349
UNet	Yes	No	0.900	0.945	#168760
Nested UNet	Yes	No	0.896	0.943	#169077
SegNet	Yes	No	0.895	0.944	#169078
UNet	No	No	0.853	0.922	#169073

Installing

To run the code of this project, you need to install the libraries listed in the requirements.txt file. You can perform the installation using this command:

pip3 install -r requirements.txt

Dependencies:

• matplotlib
• numpy
• pillow
• scikit-image
• torch
• torchvision
• tqdm

Instructions

The scripts directory contains scripts to perform the different tasks of the project.

Predictions for AIcrowd

To reproduce our submission on AIcrowd, move in the scripts folder and run:

python3 run.py

This command will create the predicted mask for each test image in the out/submission directory. The csv file for submission produced will be out/submission.csv.

Data augmentation

To create the augmented training dataset, you can run:

python3 augment_data.py

The images created will be in the data/training_augmented directory. If this directory already exists, it will overwrite the images.

Training

To train a model, you can use the train.py script:

python3 train.py

To see the different options, run python3 train.py --help.

Predicting

To create the predicted masks using a trained model, you can use the predict.py script:

python3 predict.py

To see the different options, run python3 predict.py --help.

Plotting

The pickle files created during training can be visualized using the plot_metrics.py script:

python3 plot_metrics.py --file FILE

FILE must be a pickle file.

Structure

This is the structure of the repository:

• data: contains the datasets
• docs: contains the documentation
• figs: contains the figures
• notebooks: contains the notebooks
• scripts: contains the main scripts
  • augment_data.py: create the augmented dataset
  • config.py: helpers functions to configure paths
  • plot_metrics.py: plot metrics
  • predict.py: make predictions using a trained model
  • run.py: make predictions for AIcrowd
  • train.py: train the model
• src: source code
  • models: neural network models
    • nested_unet.py: nested UNet implementation
    • segnet.py: SegNet implementation
    • unet.py: UNet implementation
  • data_augmentation.py: creation of the augmented dataset
  • datasets.py: custom dataset class for satellite images
  • loss.py: custom loss functions
  • metrics.py: score and performance functions
  • path.py: paths and archives management
  • plot_utils.py: plot utils using matplotlib
  • postprocessing.py: postprocessing functions to improve predictions
  • predicter.py: predicter class to make predictions using a trained model
  • submission.py: submission utils
  • trainer.py: trainer class to train a model

References

See references.