My solution is based on UNet with a deep supervision branch for empty mask classification.

• Model: UNet.
• Backbone: ResNet34 backbone with frozen batch-normalization.
• Preprocessing: training on random crops with (512, 512) size, inference on (768, 768) size.
• Augmentations: ShiftScaleRotate, RandomBrightnessContrast, ElasticTransform, HorizontalFlip from albumentations.
• Optimizer: Adam, batch_size=8
• Scheduler: CosineAnnealingLR
• Additional feature: the proportion of non-empty samples linearly decreased from 0.8 to 0.22 (as in train dataset) depending on the epoch. It helped to converge faster.
• Loss: 2.7 * BCE(pred_mask, gt_mask) + 0.9 * DICE(pred_mask, gt_mask) + 0.1 * BCE(pred_empty, gt_empty). Here pred_mask is the prediction of the UNet, pred_empty is the prediction of the branch for empty mask classification.
• Postprocessing: if pred_empty > 0.4 or area(pred_mask) < 800: pred_mask = empty. Parameters are selected on the validation set.
• Ensemble: averaging the 4 best checkpoints over 8 folds, horizontal flip TTA.
• Hardware: 4 x RTX 2080

make build
make run
make exec

cd scripts
bash train.sh
bash test.sh
bash ensemble.sh
bash submit.sh

• https://github.com/albu/albumentations
• https://github.com/open-mmlab/mmdetection
• https://github.com/SpaceNetChallenge/SpaceNet_Off_Nadir_Solutions/blob/master/selim_sef