Real Time Face Recognition using FaceNet and OpenCV

This is a real time face recognition project based on FaceNet and OpenCV.

Compatibility

The code is tested using Tensorflow 1.3 with GPU support under Fedora 26 with Python 2.7 and Python 3.6.

Requirements

• NumPy
• SciPy
• scikit-learn
• Pillow
• OpenCV-Python
• TensorFlow

Pre-processing

The dataset structure

face_DB/raw
├── ID1
│     ├── ID1_001.jpg
│     ├── ID1_002.jpg
│     ├── ID1_003.jpg
│     ├── ID1_004.jpg
│     └── ID1_005.jpg
├── ID2
│     ├── ID2_001.jpg
│     ├── ID2_002.jpg
│     ├── ID2_003.jpg
│     ├── ID2_004.jpg
│     └── ID2_005.jpg
├── ID3
│     ├── ID3_001.jpg
...
...

Pre-trained models

Use the Pre-trained models from davidsandberg/facenet

Align the dataset

python align_dataset_mtcnn.py <raw_img_dir> <aligned_img_dir>

Example

python align_dataset_mtcnn.py Face_db/raw Face_db/align_160

Train a classifier

python classifier.py TRAIN <aligned_img_dir> <facenet_model_path> <classifier_path>

Example

python classifier.py TRAIN Face_db/align_160/ models/20170512-110547/20170512-110547.pb models/classifier/test_classifier.pkl

Run

python camera.py <mode> <facenet_model_path> <classifier_path> --interval=5 --minsize=80

• mode
  • ONLY_DETECT: Only detects faces from the camera
  • ALL: Recognizes faces from the camera
• interval: Frame interval of each face recognition event, default value is 5
• minsize: Minimum size (height, width) of face in pixels, default value is 80

Example

python camera.py ALL models/20170512-110547/20170512-110547.pb models/classifier/test_classifier.pkl --interval=5 --minsize=80

Inspiration

• davidsandberg/facenet The following codes and files was taken from this repository:
  • faceney.py
  • detect_face.py
  • align_dataset_mtcnn.py
  • classifier.py
  • models/mtcnn/
• shanren7/real_time_face_recognition The workflow was inspired by here.