ML Project 2 - Team MNM - Option B - Road Segmentation

AI Crowd Challenge link

For this project task, we had to create a model to segment roads in satellite images, i.e assign a label road=1, background=0 to each pixel.

The dataset is available in this git aswell as on the CrowdAI.

Our best model achieved an F1-score of 90.7% and an accuracy of 95%. Final submission: Submission #169329

Our Report

How to reproduce our best submission

Clone this repo and follow the setup and run steps below!

Environment Setup

Run the following commands to create an appropriate python environment and install all required libraries.

conda create -y -n ml_roadseg python=3.9.7 scipy pandas numpy matplotlib
conda activate ml_roadseg
pip install Pillow
pip install opencv-python
conda install -y pytorch torchvision torchaudio -c pytorch

Running the code

# Activate python environment
conda activate ml_roadseg

# Run preprocessing/data augmentation
python run.py prepro

# Runs model training and saves model
python run.py train_model

# Loads trained model and runs predictions on test set
python run.py predict_test

# Read predicted labels and write them to the .csv submission format
python run.py write_sub

Authors

Baldwin Nicolas - chabala98

Leidi Mauro - MauroLeidi

Roust Michael - michaelroust