By Pouya Samangouei*, Mahyar Najibi*, Larry Davis, Rama Chellappa

* contributed equally.

The paper can be found here.

This Python implementation is based on Faster-RCNN.

Face-MagNet is released under the Apache 2 License (refer to the LICENSE file for details).

@inproceedings{wacv18facemagnet,
    Author = {Pouya Samangouei and Mahyar Najibi and Larry Davis and Rama Chellappa},
    Title = {Face-MagNet: Magnifying Feature Maps to Detect Small Faces},
    Booktitle = {IEEE Winter Conf. on Applications of Computer Vision ({WACV})},
    Year = {2018}
}

• Download face-magnet.

cd <project-parent-dir>
git clone --recursive https://github.com/po0ya/face-magnet
FM_ROOT=`pwd`/face-magnet

• Install python requirements.

  cd $FM_ROOT
  pip install -r requirements.txt
  

• Make lib:

  cd $FM_ROOT/lib
  make
  

• Install nccl from https://developer.nvidia.com/nccl for multi-gpu training.

• Build caffe.

  cd caffe-facemagnet
  # Set build flags e.g., edit caffe-facemagnet/Makefile.config appropriately.
  make -j8
  make pycaffe
  

• Download the WIDER-Face dataset.

  WIDER_PATH=<path-to-wider-face>
  

• Perpare the data directory:

  • Download imagenet pretrained weights using scripts from Faster-RCNN:

    ./data/scripts/fetch_imagenet_models.sh
    

  • Create a symlink to WIDER-Face path:

    ln -s $WIDER_PATH data/wider
    

    • (optional) Copy imglist.csv files into WIDER_PATH:

    cp -r data/WIDER_imglists/* $WIDER_PATH
    

    These csv files are generated with matlab/wider_csv.m and is provided in case matlab is not available. If it is, the code will generate these automatically.

• Create a or link a debug directory for output logs.

  mkdir debug
  

  or

  ln -s <path-to-debug-dir> debug
  

./experiments/scripts/train_min800.sh <model-name> <gpu-ids>

• <model-name> can be:

  • facemagnet: Face-MagNet model for the 8th row of Table 1 and the final results.
  • sizesplit: SizeSplit model for the 7th row of Table 1.
  • skipface: SkipFace model for the 6th row of Table 1.
  • context: Context model for the 5th row of Table 1.
  • default: For the first row of Table 1.
  • facemagnet_nc, sizesplit_nc, and skipface_nc for the models without context pooling.

• <gpu-ids> is a string of comma separated GPU ids.

• To see the training progress tail -f <log-path> where <log-path> is shown after running the above script.

For example, for facemagnet run:

./experiments/scripts/train_min800.sh facemagnet 0,1

To train Face-MagNet on two GPUs.

For single GPU training please use ./tools/train_net.py.

 python ./tools/train_net.py --gpu <gpu-id> \
  --solver models/wider/VGG16/solvers/<model-name>.prototxt \
  --weights data/imagenet_models/VGG16.v2.caffemodel \
  --imdb wider_train \
  --iters 76000 \
  --cfg experiments/cfgs/min800.yml \
  ${EXTRA_ARGS}

To detect faces on a single image with a trained model:

./experiments/scripts/test_single.sh <model-name> <image-path> <args>

• <model-name> is the same as the training section.
• <image-path> is the path to the image
• args can be the following optional arguments:
  • --min_size <int> Overrides cfg.TEST.SCALES so that the minimum dimension of the image is scaled to this value.
  • --max_size <int> Overrides cfg.TEST.MAX_SIZE so that the maximum dimension is not greater than its value.
  • --pyramid Performs the detection three times on scales [0.5, 1, 2] of the image and combines the result.

For example to detect the faces in images of data/demo:

./experiments/scripts/test_single.sh facemagnet data/demo/demo0.jpg --pyramid

• To produce the results of each row of Table 1, use:

  ./experiments/scripts/test_min800.sh <model-name>
  

• To perform the benchmark test on WIDER-Face:

  ./experiments/scripts/test_wider.sh
  

• To test the fddb and pascal datasets:

  • Pascal image list can be found in data/pascal_imglist.csv
  • Download FDDB and link it to data/fddb.
  • Prepare Pascal Faces and put it under data/pascal.
  • To get the detections:

    ./experiments/scripts/test_base.sh VGG16 {fddb|pascal} <model-name> 38000 ./experiments/cfgs/min800.yml
    

  • To produce eval plots see tools/marcopede-face_eval.
  • Take a look at and use ./experiments/scripts/plot_fddb.sh for FDDB.

• The results are reported using two NVIDIA P6000 GPU with 24GB GPU memory. Big images need a higher memory because of the sizes of the convolutional feature maps, therefore if you're facing memory issues try setting cfg.TEST.MAX_SIZE to a smaller number.

• Uploading the models.