gf0507033 / FaceLib

Face Detection & Age Gender & Expression & Recognition

FaceLib:

• use for Detection, Facial Expression, Age & Gender Estimation and Recognition with PyTorch
• this repository works with CPU and GPU(Cuda)

New Version (coming next month)

• improve performance
• how can you train it with new data?
• add some new features
  • add some good default argument for functions
  • download weight files into codes
  • create pip package
  • some example in jupyter notebooks

Installation

• Clone and install with this command:

  pip install git+https://github.com/sajjjadayobi/FaceLib.git

• Prerequisites

  • Python 3.6+
    • PyTorch 1.4+
  • Torchvision 0.4.0+
    • OpenCV 2.0+
  • requirements.txt

1. Face Detection: RetinaFace

• you can use these backbone networks: Resnet50, mobilenet, slim, RFB
• for more details, you can use the documentation
• The following example illustrates the ease of use of this package:

 from facelib import FaceDetector
 detector = FaceDetector(name='mobilenet', weight_path='mobilenet.pth', device='cpu')
 boxes, scores, landmarks = detector.detect_faces(image)

• downlaod weight of network from google drive RetinaFace
• you can cheche code and change it Face Detector
• based on this repository RetinaFace

WiderFace Validation Performance on a single scale When using Mobilenet for backbone

Style	easy	medium	hard
Pytorch (same parameter with Mxnet)	88.67%	87.09%	80.99%
Pytorch (original image scale)	90.70%	88.16%	73.82%
Mxnet(original image scale)	89.58%	87.11%	69.12%

2. Face Alignment: Similar Transformation

• you can use this module like this:

   from facelib import FaceDetector
   detector = FaceDetector(name='mobilenet', weight_path='mobilenet.pth', device='cuda')
   faces, boxes, scores, landmarks = detector.detect_align(image)

• or run on webcam and shows the result on the image

  python Retinaface/from_camera.py

• detect_image() instead detect_faces()

• for more details read detect_image function documentation

• let's see a few examples

Original	Aligned & Resized	Original	Aligned & Resized

3. Age & Gender Estimation:

• I used UTKFace DataSet for Age & Gender Estimation
• you can use these backbone networks: full, tiny
• you can use this module like this:

   from facelib import FaceDetector
   from facelib import AgeGenderEstimator

   face_detector = FaceDetector(name='mobilenet', weight_path='mobilenet.pth', device='cuda')
   age_gender_detector = AgeGenderEstimator(name='full', weight_path='ShufflenetFull.pth', device='cuda')

   faces, boxes, scores, landmarks = face_detector.detect_align(image)
   genders, ages = age_gender_detector.detect(faces)
   print(genders, ages)

• or run on webcam and shows the result on the image

  python AgeGender/from_camera.py

• downlaod weight of network from google drive ShufleNet

4. Facial Expression Recognition:

• Facial Expression Recognition using Residual Masking Network
• face size must be (224, 224), you can fix it in FaceDetector init function with face_size=(224, 224)

  from facelib import FaceDetector
  from facelib import EmotionDetector

  face_detector = FaceDetector(name='mobilenet', weight_path='mobilenet.pth', face_size=(224, 224))
  emotion_detector = EmotionDetector(name='resnet34', weight_path='resnet34.pth', device='cuda')

  faces, boxes, scores, landmarks = face_detector.detect_align(image)
  list_of_emotions, probab = emotion_detector.detect_emotion(faces)
  print(list_of_emotions)

• or run on webcam and shows the result on the image

  python FacialExpression/from_camera.py

• downlaod weight of network from google drive Expression

• like this image:

5. Face Recognition: InsightFace

• This module is a reimplementation of Arcface(paper), or Insightface(Github)
• For models, including the PyTorch implementation of the backbone modules of IR-SE50 and MobileFacenet

Pretrained Models & Performance

• IR-SE50

LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	calfw(%)	cplfw(%)	vgg2_fp(%)
0.9952	0.9962	0.9504	0.9622	0.9557	0.9107	0.9386

• Mobilefacenet

LFW(%)	CFP-FF(%)	CFP-FP(%)	AgeDB-30(%)	calfw(%)	cplfw(%)	vgg2_fp(%)
0.9918	0.9891	0.8986	0.9347	0.9402	0.866	0.9100

Prepare the Facebank (For testing over camera or video)

• Provide the face images your want to detect in the data/face_bank folder, and guarantee it have a structure like following:

  data/facebank/
          ---> person_1/
              ---> img_1.jpg
          ---> person_2/
              ---> img_1.jpg
              ---> img_2.jpg
  

• you can save a preson with 3 ways:

  • use python add_face_from_camera.py -n NAME
    • use python add_face_from_dir.py -n NAME
    • or add faces manually (just face of person not image of a person)

• you can use this module like this for camera verification:

    python camera_verify.py -u update -m True
  

  • u argument: update FaceBank if add a new person
  • m argument: use Mobilenet for backbone

• and use into your code:

    from facelib import get_config
    from facelib import FaceRecognizer
    from facelib import update_facebank, load_facebank, special_draw
    from facelib import FaceDetector


    conf = get_config(training=False)
    detector = FaceDetector(name='mobilenet', weight_path='mobilenet.pth', device=conf.device)
    conf.use_mobilfacenet = True or False
    face_rec = FaceRecognizer(conf, inference=True)
    face_rec.model.eval()
    
    # set True when you add someone new 
    update_facebank_for_add_new_person = False
    if update_facebank_for_add_new_person:
        targets, names = update_facebank(conf, face_rec.model, detector)
    else:
        targets, names = load_facebank(conf)

    faces, boxes, scores, landmarks = detector.detect_align(image)
    results, score = face_rec.infer(conf, faces, targets)
    for idx, bbox in enumerate(boxes):
        special_draw(image, bbox, landmarks[idx], names[results[idx]+1], score[idx])

• downlaod weight of network from google drive InsightFace
• example of run this code:

Reference:

• InsightFace
• RetinaFace
• Facial Expression

Citation:

   - Author : Sajjad Ayoubi
   - Title : FaceLib
   - Year = 2020