Assessing out-of-domain generalization for robust building damage detection

This repository contains code for the paper "Assessing out-of-domain generalization for robust building damage detection" by Vitus Benson and Alexander Ecker.

We use the Dual-HRNet from https://github.com/DIUx-xView/xView2_fifth_place and the xView2-Score from https://github.com/DIUx-xView/xView2_scoring.

Setup

Get yourself a copy of the xBD dataset from https://xView2.org

Install Anaconda or Miniconda.

Run

bash -i setup.sh /path/to/xbd/

This will install a conda environment robust-bdd and prepare the xBD data for the two-stream ResNet50.

Download https://1drv.ms/u/s!Aus8VCZ_C_33dYBMemi9xOUFR0w with your web-browser and move the file to robust-bdd/dualhrnet.

These are the ImageNet-pretrained weights for HRNetv2 coming from https://github.com/HRNet/HRNet-Image-Classification.

Use

Note: All of the following requires a GPU. Our experiments were performed either with 2 GTX1080 or with 1 RTX 2080TI.

You can re-train the models presented in the Paper by running:

bash train.sh path/to/xbd/ <Table number> <twostream-resnet50 or dualhrnet> <plain or multi>

You can reproduce the inference by the models run in the Paper with:

bash test.sh path/to/xbd/ <Table number> <twostream-resnet50 or dualhrnet> <plain or swa or multiadabn or multiadabnswa>

Weights

Pre-trained weights are automatically downloaded when using the above test.sh script.

If you want to manually download them, find them under:

Benson, Vitus, 2020, "Replication Data for: Assessing out-of-domain generalization for robust building damage detection", https://doi.org/10.25625/TRTODX, Göttingen Research Online / Data, V2

Citation

If you use this work, please cite:

@article{Benson2020,
    Author = {Vitus Benson and Alexander S. Ecker},
    Title = {Assessing out-of-domain generalization for robust building damage detection},
    Year = {2020},
    journal = {NeurIPS 2020 Workshop on Artificial Intelligence for Humanitarian Assistance and Disaster Response (AI+HADR 2020)},
    URL = {https://arxiv.org/abs/2011.10328}
}