Rooftop Detection Challenge

Installation

• Create virtual environment (python <=3.6) virtualenv venv
• Activate environment (bash) source venv/bin/activate
• Install requirements pip install -r requirements.txt

Start training

• python segmentation.py

I used

• Pretrained model from https://github.com/qubvel/segmentation_models
• Additional data from https://www.airs-dataset.com/ (I did not have the time to take full advantage of it). Data is too large to put here(~17gb) and possibly proprietary
• https://dataturks.com/ for annotation of validation images (I did not train on them, just for easier model evaluation)

Next steps

• make more use of large christchurch dataset (prepare/augment data better)
• Try different models/losses (dice loss seems better from intuition since it aligns with IoU score which is very informative in my opinion)