This solution gets 0.9945 dice, which is within 0.25% from current top, but still places in the middle of the leaderboard (competition requires pixel-perfect segmentation).
This is mainly for-fun solution, without any special tricks (no ensembling, no data augmentation), with very little hyperparameter tuning.
Architecture: slightly customized UNet (so that it can process images in native 1918x1280 resolution)
Framework: PyTorch
- Extract competition data in
input/ - Preprocess data with
python dataset.py(20 min) - Train to convergence with
python train.py(12+ hours) - If necessary, generate submission with
python make_submission.py
MIT