Our human meshes, background images, and yolov2 model file can be downloaded from: https://mega.nz/file/pZtUCKza#6AF3AkIYxiWXysqoo78nbjKoTCos6-PwU_UBaSntIA8
After extracting the .zip file, your directory should contain
./data/background
./data/meshes
./data/test_background
./data/yolov2
PyTorch 1.8.0 and Torchvision 0.8.1 (tested in python 3.8):
pip install torch torchvision
PyTorch3D v0.2.5:
pip install 'git+https://github.com/facebookresearch/[email protected]'
To train/test on Faster R-CNN:
git clone https://github.com/potterhsu/easy-faster-rcnn.pytorch.git faster_rcnn
cd faster_rcnn
python support/setup.py develop
Faster R-CNN pretrained checkpoint can be found here: https://github.com/potterhsu/easy-faster-rcnn.pytorch
Move the model-180000.pth checkpoint to faster_rcnn/model-180000.pth.
To train an adversarial patch:
python train.py --mesh_dir=data/meshes --epochs=100 --num_bgs=1024 --num_test_bgs=1024 --batch_size=12 --num_angles_train=1 --angle_range_train=0 --num_angles_test=21 --angle_range_test=10 --idx=idx/chest_legs1.idx --detector=yolov2 --patch_dir=example_logos/fasterrcnn_chest_legs
During training, the script will save the adversarial texture atlas and face indices to patch_save.pt and idx_save.pt respectively. The --idx argument specifies the indices of the patch. To generate new .idx files, see the Blender script face_sampler.py.
To test a patch:
python train.py --test_only --mesh_dir=data/humans --num_test_bgs=1024 --num_angles_test=21 --angle_range_test=10 --patch_dir=my_patch --detector=yolov2
The --patch_dir argument specifies a folder that contains the patch_save.pt and idx_save.pt files.