README.md

Face Detection

Face detection with different models. Use -h flag for help. Requirements must be installed beforehand. Note if the webcam mode is not working, make sure the terminal has permission to access the webcam.

blazeface face detection

python detect_face_blazeface.py                     # webcam mode default
python detect_face_blazeface.py -i PATH_TO_IMG/VID  # image or video mode

mtcnn face detection

python detect_face_mtcnn.py                     # webcam mode default
python detect_face_mtcnn.py -i PATH_TO_IMG/VID  # image or video mode

cv2 DNN face detection

python detect_face_opencv_dnn.py                     # webcam mode default
python detect_face_opencv_dnn.py -i PATH_TO_IMG/VID  # image or video mode

OpenVINO face detection

python detect_face_openvino.py                     # webcam mode default
python detect_face_openvino.py -i PATH_TO_IMG/VID  # image or video mode

YOLOv5 face detection

python detect_face_yolov5_face.py                     # webcam mode
python detect_face_yolov5_face.py -i PATH_TO_IMG/VID  # image or video mode

Face Detection model evaluation

Evaluation on WIDERFACE dataset

Install pycocotools

pycocotools is required for evaluation script.

Inside a virtual env.

pip install Cython
# make and install pycocotools
git clone https://github.com/cocodataset/cocoapi
cd coco/PythonAPI
make
pip install .

Download WIDER_val split and run evaluation

Download wider_face_split and WIDER_val from http://shuoyang1213.me/WIDERFACE/ and place them inside eval directory.

# by default run wider_face_val_split with opencv face detection
PYTHONPATH=$PYTHONPATH:./ python eval/eval_face_detector.py eval/wider_face_split/wider_face_val_bbx_gt.txt eval/WIDER_val/images/
# run wider_face_val_split with yolov5s face detection
PYTHONPATH=$PYTHONPATH:./ python eval/eval_face_detector.py eval1/wider_face_split/wider_face_val_bbx_gt.txt eval1/WIDER_val/images/ --model_type yolov5_face --model weights/yolov5s/yolov5s-face.pt

Face Extraction from Dataset

Extract face images in bulk from an image dataset.

Face Extraction from imdb-wiki dataset

Download the imdb-wiki data from IMDB-WIKI and place it inside the face_detection_and_extraction/data directory.

python face_extraction/extract_and_clean_imdb_wiki_faces.py -d {data/imdb, data/wiki}

Face Extraction from dataset

python face_extraction/extract_faces_from_dataset.py --rd SRC_DATA_DIR --td TARGET_DATA_DIR
python face_extraction/extract_faces_from_dataset.py --rd data --td extracted_faces  # using existing imgs in data

Raw Data Organization:

SRC_DATA_DIR
            |_ class_1
                     |_ img1/vid1
                     |_ img2/vid2
                     |_ ....
            |_ class_2
                     |_ img1/vid1
                     |_ img2/vid2
                     |_ ....
            ...

Face Extraction and Labelling from dataset

python face_extraction/extract_and_label_faces_from_dataset.py --rd SRC_DATA_DIR --td TARGET_DATA_PATH

Annotation data is also saved in the following pickle format:

{
  "media_id": "0123456789",
  "media_url": "http://example/0123456789",
  "class_name": "class_1",
  "frame_info": [{
      "frame_num": 0,
      "time_sec": 0,
      "face_ids": [0, 1],
      "face_bboxes": [
        [1, 2, 3, 4],
        [5, 6, 7, 8]
      ],
      "confs": [0.6, 0.8],
      "ages": ["20-50", "20-50"],
      "genders": ["male", "female"],
      "normed_embedding": ["[0.1, ..., 0.2]", "[0.3, ..., 0.4]"]
    },
    {
      "frame_num": 1,
      "time_sec": 1,
      "face_ids": [0, 1],
      "face_bboxes": [
        [1, 2, 3, 4],
        [5, 6, 7, 8]
      ],
      "confs": [0.8, 0.9],
      "ages": ["0-5", "20-50"],
      "genders": ["male", "female"],
      "normed_embedding": ["[0.1, ..., 0.2]", "[0.3, ..., 0.4]"]
    }
  ]
}

For Developers

Running Tests

Install requirements

pip install requirements/dev.txt

Run tests from inside the face_detection_and_extraction directory

python -m pytest tests                 # run pytest
python -m pytest --cov=modules tests/  # run pytest with coverage