Official Pytorch implementation for our AAAI 2023 paper HOTCOLD Block: Fooling Thermal Infrared Detectors with a Novel Wearable Design.

• python 3.9
• Pytorch 1.10
• At least 1x12GB NVIDIA GPU

git clone https://github.com/weihui1308/HOTCOLDBlock
cd HOTCOLDBlock-main
pip install -r requirements.txt

1. Download the complete FLIR ADAS Dataset and convert its annotation format to the YOLO format.
2. Filter out instances of "person" from the dataset, and keep only those instances with a height greater than 120 pixels.
3. We have placed the conversion script json2yolo.py in the dataset folder.
4. Put the obtained dataset in YOLO format in the "dataset/FLIR_ADAS" folder.

1. Download the YOLOv5 pre-trained model. In this work, we use the YOLOv5s.pt.
2. Fine-tune the pre-trained YOLOv5 model on the "dataset/FLIR_ADAS". You can download the model of our training at Google Drive.

Once you have setup your path, you can run an experiment like so:

python main.py --epochs 5 

The terminal will print the gbest_position and gbest_value.

If you find this repository useful, please consider citing our paper:

@inproceedings{wei2023hotcold,
  title={HOTCOLD Block: Fooling Thermal Infrared Detectors with a Novel Wearable Design},
  author={Hui Wei and Zhixiang Wang and Xuemei Jia and Yinqiang Zheng and Hao Tang and Shin'ichi Satoh and Zheng Wang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2023}
}

We would like to acknowledge the YOLOv5 open-source library (https://github.com/ultralytics/yolov5). YOLOv5 is a powerful object detection algorithm that has greatly facilitated our development efforts. We are grateful to the developers and contributors of YOLOv5 for making their work available to the community.