3rd place solution for xView2 Damage Assessment Challenge
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.
Please check out our paper that describes the approach in more details: https://arxiv.org/abs/2111.00508
Approach in a nutshell
- 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.
Training
- Install dependencies from
requirements.txt - Follow
train.sh
Inference
For inference using pre-trained models please download full archive from Releases tab and run predict_37_weighted.py script.
License
MIT