This repository is based on the marvis YOLOv2 inplementation: https://github.com/marvis/pytorch-yolo2

This work corresponds to the following paper: https://arxiv.org/abs/1904.08653:

@inproceedings{thysvanranst2019,
    title={Fooling automated surveillance cameras: adversarial patches to attack person detection},
    author={Thys, Simen and Van Ranst, Wiebe and Goedem\'e, Toon},
    booktitle={CVPRW: Workshop on The Bright and Dark Sides of Computer Vision: Challenges and Opportunities for Privacy and Security},
    year={2019}
}

If you use this work, please cite this paper.

We use Python 3.6. Make sure that you have a working implementation of PyTorch installed, to do this see: https://pytorch.org/ To visualise progress we use tensorboardX which can be installed using pip:

pip install tensorboardX tensorboard

No installation is necessary, you can simply run the python code straight from this directory.

Make sure you have the YOLOv2 MS COCO weights:

mkdir weights; curl https://pjreddie.com/media/files/yolov2.weights -o weights/yolo.weights

Get the INRIA dataset:

curl ftp://ftp.inrialpes.fr/pub/lear/douze/data/INRIAPerson.tar -o inria.tar
tar xf inria.tar
mv INRIAPerson inria
cp -r yolo-labels inria/Train/pos/

patch_config.py contains configuration of different experiments. You can design your own experiment by inheriting from the base BaseConfig class or an existing experiment. ReproducePaperObj reproduces the patch that minimizes object score from the paper (With a lower batch size to fit on a desktop GPU).

You can generate this patch by running:

python train_patch.py paper_obj