A VGG-Face CNN descriptor implemented in PyTorch.

The VGG-Face CNN descriptors are computed using [1] authors' CNN implementation, based on the VGG-Very-Deep-16 CNN architecture (see [1]), and are evaluated on the Labeled Faces in the Wild [2] and the YouTube Faces [3] datasets. The provided models can be used for non-commercial research purposes under Creative Commons Attribution License.

Step 0: Setup and activate a virtual environment (Python 3.3+), as follows:

$ python3 -m venv venv
$ source venv/bin/activate
$ pip install --upgrade pip
$ pip install -r requirements.txt
$ source venv/bin/activate

Step 1: Convert the original pre-trained LuaTorch weights into PyTorch VGGFace weights and save them at models/vggface.pth by running the following script:

$ python3 convert_weights.py 

After this step, models directory should be as follows:

models/
├── vggface.pth
├── vgg_face_torch
│   └── VGG_FACE.t7
└── vgg_face_torch.tar.gz

Step 2: Run demo script:

$ python3 demo.py 

By default, image data/rm.png is used (use a different image using --img=<image_file>) -- the output should be as follows:

Predicted id: Rooney_Mara (probability: 0.984787964730149)

Get VGGFace sub-model

Use the following script for getting a sub-model of VGGFace, e.g., only the convolutional sub-network, and/or rename its layers:

$ python3 get_submodel.py -h
usage: Get (sub)model of VGGFace model [-h] [--model MODEL] [--output OUTPUT]

optional arguments:
  -h, --help       show this help message and exit
  --model MODEL    input VGGFace model file
  --output OUTPUT  output VGGFace (sub)model file

References

[1] Parkhi, Omkar M., Andrea Vedaldi, and Andrew Zisserman. "Deep face recognition." BMVC. Vol. 1. No. 3. 2015.

[2] G. B. Huang, M. Ramesh, T. Berg, E. Learned-Miller Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, 2007.

[3] L. Wolf, T. Hassner, I. Maoz "Face Recognition in Unconstrained Videos with Matched Background Similarity." Computer Vision and Pattern Recognition (CVPR), 2011.