Incidents Dataset

See the following pages for more details:

• Project page: http://incidentsdataset.csail.mit.edu/ or https://ethanweber.me/IncidentsDataset.
• ECCV 2020 Paper "Detecting natural disasters, damage, and incidents in the wild" here.
• Extended Paper "Incidents1M: a large-scale dataset of images with natural disasters, damage, and incidents" here.

Obtain the data

Please fill out this form and then email/notify incidentsdataset@googlegroups.com to request the data.

The data structure is in JSON with URLs and labels. The files are in the following form:

# single-label multi-class (ECCV 2020 version):
eccv_train.json
eccv_val.json

# multi-label multi-class (latest version):
multi_label_train.json
multi_label_val.json

1. Download chosen JSON files and move to the data folder.

2. Look at VisualizeDataset.ipynb to see the composition of the dataset files.

3. Download the images at the URLs specified in the JSON files.

4. Take note of image download location. This is param --images_path in parser.py.

Setup environment

git clone https://github.com/ethanweber/IncidentsDataset
cd IncidentsDataset

conda create -n incidents python=3.8.2
conda activate incidents
pip install -r requirements.txt

Using the Incident Model

1. Download pretrained weights here. Place desired files in the pretrained_weights folder. Note that these take the following structure:

   # run this script to download everything
   python run_download_weights.py
   
   # pretrained weights with Places 365
   resnet18_places365.pth.tar
   resnet50_places365.pth.tar
   
   # ECCV baseline model weights
   eccv_baseline_model_trunk.pth.tar
   eccv_baseline_model_incident.pth.tar
   eccv_baseline_model_place.pth.tar
   
   # ECCV final model weights
   eccv_final_model_trunk.pth.tar
   eccv_final_model_incident.pth.tar
   eccv_final_model_place.pth.tar
   
   # multi-label final model weights
   multi_label_final_model_trunk.pth.tar
   multi_label_final_model_incident.pth.tar
   multi_label_final_model_place.pth.tar
   

2. Run inference with the model with RunModel.ipynb.

3. Compute mAP and report numbers.

   # test the model on the validation set
   python run_model.py \
       --config=configs/eccv_final_model \
       --mode=val \
       --checkpoint_path=pretrained_weights \
       --images_path=/path/to/downloaded/images/folder/
   

4. Train a model.

   # train the model
   python run_model.py \
       --config=configs/eccv_final_model \
       --mode=train \
       --checkpoint_path=runs/eccv_final_model
   
   # visualize tensorboard
   tensorboard --samples_per_plugin scalars=100,images=10 --port 8880 --bind_all --logdir runs/eccv_final_model
   

   See the configs/ folder for more details.

Citation

If you find this work helpful for your research, please consider citing our paper:

@InProceedings{weber2020eccv,
  title={Detecting natural disasters, damage, and incidents in the wild},
  author={Weber, Ethan and Marzo, Nuria and Papadopoulos, Dim P. and Biswas, Aritro and Lapedriza, Agata and Ofli, Ferda and Imran, Muhammad and Torralba, Antonio},
  booktitle={The European Conference on Computer Vision (ECCV)},
  month = {August},
  year={2020}
}

License

This work is licensed with the MIT License. See LICENSE for details.

Acknowledgements

This work is supported by the CSAIL-QCRI collaboration project and RTI2018-095232-B-C22 grant from the Spanish Ministry of Science, Innovation and Universities.