Awesome

Siamese Attention U-Net for Multi-Class Change Detection

1st place solution for the EARTHVISION 2021 DynamicEarthNet Challenge - Weakly-Supervised Multi-Class Change Detection Track at CVPRW 2021
[Challenge Site] [Presentation Slides] [Presentation Video] [Dataset]

Overview

This work introduces a pixel-wise change detection network named Siamese Attention U-Net that incorporates attention mechanisms in the Siamese U-Net architecture. Experiments show the architectural change alongside training strategies such as semi-supervised learning produce more robust models.

Results

1. Attention block
   

When training the model with Jaccard loss, attending the up-sampled features improves mean IoU scores whereas attending skipped features - originally proposed in Attention UNet - do not. The improvement implies a need for prioritizing information in the coarser resolution features.

2. Loss function

Jaccard loss and Dice loss optimize for different metrics, producing slightly varying results. Ensembling models trained exclusively on each loss function improves scores.

3. Semi-supervised learning

Creating hard pseudo labels for the public validation and test dataset, then retraining the model alongside the original training dataset improves scores regardless of loss function. The ensemble of models trained on each loss function is submitted to the public test benchmark.

Usage

Dependencies

• gdal
• numpy
• pandas
• pillow
• pytorch
• pyyaml
• torchvision
• tqdm

Downloading and Preprocessing Data

# assumes all data is unzipped to ./earthvision2021/data/source/
cd ./setup/
# convert data and organize directories
bash setup.bash

cd ./earthvision2021/preprocessing/
# create csv file that determines crop regions for training
python crop.py
mv train.csv ../data/train/random_128_binary_attempts128.csv

Training

cd ./earthvision2021/
# train model according to configurations in ./earthvision2021/config_train.yml
python train.py

Predicting

cd ./earthvision2021/
# predict using model according to configurations in ./earthvision2021/config_predict.yml
python predict.py

Pseudo Labelling

cd ./earthvision2021/postprocessing/
# binarizing prediction results to create hard pseudo labels
python label_binarize.py --in_dir ${prediction_dir}
cd ../preprocessing/
# create new csv file of crop regions
python crop.py --subset_file ${filename_text_file}
mv train.csv ../data/train/${csv_file}
# change "aoi_file" in ../config_train.yml to ${csv_file} and retrain

Postprocessing

cd ./earthvision2021/postprocessing/
# ensemble multiple predictions results
python ensemble.py --in_dir_list ${prediction_dir} ...
# binarize prediction results
python binarize.py --in_dir ${prediction_dir}

Citation

If you find this work useful in your research or publication, please cite this work:

@inproceedings{cummings2022_sia,
  author={Cummings, Sol and Kondmann, Lukas and Zhu, Xiao Xiang},
  booktitle={IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium},
  title={Siamese Attention U-Net for Multi-Class Change Detection},
  year={2022},
  pages={211-214},
  doi={10.1109/IGARSS46834.2022.9884834}
}