Before generating the attack images, the input should be processed by face_landmark_detection.py.
Args:
- input folder
- wildcard (ex. if there's a file structure)
This will generate the images that can run through MatConvNet.
The next step is to run align_vgg_pose which can be run from generate.m if digital_dodging or equivalent is replaced by a call to that function passing in the input path, wildcard and output path.
After runnning the preprocess, generate.m will create the attack images (for now, digital dodging as the only mode).
A GTX2070 card was used which has 8 GB of VRAM, but MatConvNet threw out of memory errors on some images (all were 224x224).
Either better hardware or scaled down images should fix this issue.
Accessorize-to-a-crime has all necessary dependencies and is the backbone of this little tool