Eugene Khvedchenya, February 2020

This repository contains source code for my solution to xView2 challenge. My solution was scored second (0.803) on public LB and third (0.805) on private hold-out dataset.

• Ensemble of semantic segmentation models.
• Trained with weighted CE to address class imbalance.
• Heavy augmentations to prevent over-fitting and increase robustness to misalignment of pre- and post- images.
• Shared encoder for pre- and post- images. Extracted feature are concatenated and sent to decoder.
• Bunch of encoders (ResNets, Densenets, EfficientNets) and two decoders: Unet and FPN.
• 1 round of Pseudolabeling
• Ensemble using weighted averaging. Weights optimized for every model on corresponding validation data.

• Install dependencies from requirements.txt
• Follow train.sh

For inference using pre-trained models please download full archive from Releases tab and run predict_37_weighted.py script.

MIT