WIT-UAS: A Wildland-fire Infrared Thermal Dataset to Detect Crew Assets From Aerial Views

About

• This dataset contains bounding box annotated Wildland-fire Infrared Thermal (WIT) images for crew assets detection with Unmanned Aerial Systems (UAS). It is captured during prescribed burns. The associated paper can be found here. The labeled thermal data can be downloaded here. For access to the ROS bags, please email [email protected].

• Available Labels

Test Object Detection

Clone the repo locally:

git clone --recursive https://github.com/castacks/WIT-UAS-Dataset.git

Environment Setup

There are 2 options:

• Docker (recommended):

  1. Install Docker and nvidia-docker2

  2. Run the pre-built docker image (automatically pulls when missing):

     ./scripts/run.sh

  3. Attach to the running container:

     docker attach wit-uas-dataset

• Conda/Mamba environment manager:

  1. Install Anaconda or Mamba

  2. Create environment for WIT:

     mamba env create -f environment.yaml

  3. Activate WIT:

     mamba activate wit-uas

Visualization

• Visualization using wandb is optional, but when needed:

  1. Register an account on wandb.ai

  2. Login to your wandb account locally:

     wandb login

Send an email to Mukai (Tom Notch) Yu if you are not invited to team cmu-ri-wildfire.

Run training on WIT dataset

To train the models: ./model/yolo/train.py or ./model/ssd/train.py

You may need to adjust the batch size according to your GPU memory by using the --batch-size argument

How To Contribute

1. Clone the repo locally

   git clone --recursive https://github.com/castacks/WIT-UAS-Dataset.git

2. Create a new branch for your work:

   git checkout -b <branch-name>

3. Setup required development environment

   ./scripts/setup.sh

Citation

If you find this repository useful for your work, please cite the following paper:

@INPROCEEDINGS{jong2023wit,
  author={Jong, Andrew and Yu, Mukai and Dhrafani, Devansh and Kailas, Siva and Moon, Brady and Sycara, Katia and Scherer, Sebastian},
  booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={WIT-UAS: A Wildland-Fire Infrared Thermal Dataset to Detect Crew Assets from Aerial Views}, 
  year={2023},
  pages={11464-11471},
  url = {https://arxiv.org/pdf/2312.09159},
  doi={10.1109/IROS55552.2023.10341683}
}