Skip to content

A collection of deep learning frameworks ported to Keras for face analysis.

Notifications You must be signed in to change notification settings

shaoanlu/face_toolbox_keras

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

55 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

face-toolbox-keras

A collection of deep learning frameworks ported to Keras for face detection, face segmentation, face parsing, iris detection, and face verification.

Descriptions

This repository contains deep learning frameworks that we collected and ported to Keras. We wrapped those models into separate modules that aim to provide their functionality to users within 3 lines of code.

*Each module follows the license of their source repo. Notice that some models were trained on dataset with non-commercial license.

Usage

Open In Colab (Please run pip install keras==2.2.4 before initializaing models.)

This colab demo requires a GPU instance. It demonstrates all face analysis functionalities above.

1. Face detection

models.detector.face_detector.FaceAlignmentDetector(fd_weights_path=..., lmd_weights_path=..., fd_type="s3fd")

Arguments

  • fd_weights_path: A string. Path to weights file of the face detector model.
  • lmd_weights_path: A string. Path to weights file of the landmarks detector model.
  • fd_type: A string. Face detector backbone model of either s3fd or mtcnn.

Example

from models.detector import face_detector

im = cv2.imread(PATH_TO_IMAGE)[..., ::-1]
fd = face_detector.FaceAlignmentDetector()
bboxes = fd.detect_face(im, with_landmarks=False)

2. Face landmarks detection

The default model is 2DFAN-4. Lite models of 2DFAN-1 and 2DFAN-2 are also provided.

GPU 2DFAN-1 2DFAN-2 2DFAN-4
K80 74.3ms 92.2ms 133ms

Example

from models.detector import face_detector

im = cv2.imread(PATH_TO_IMAGE)[..., ::-1]
fd = face_detector.FaceAlignmentDetector()
bboxes, landmarks = fd.detect_face(im, with_landmarks=True)

3. Face parsing

models.parser.face_parser.FaceParser(path_bisenet_weights=...)

Arguments

  • path_bisenet_weights: A string. Path to weights file of the model.

Example

from models.parser import face_parser

im = cv2.imread(PATH_TO_IMAGE)[..., ::-1]
fp = face_parser.FaceParser()
# fp.set_detector(fd) # fd = face_detector.FaceAlignmentDetector()
parsing_map = fp.parse_face(im, bounding_box=None, with_detection=False)

4. Eye region landmarks detection

models.detector.iris_detector.IrisDetector()

Faster face detection using MTCNN can be found in this repo.

Example

from models.detector import iris_detector

im = cv2.imread(PATH_TO_IMAGE)[..., ::-1]
idet = iris_detector.IrisDetector()
idet.set_detector(fd) # fd = face_detector.FaceAlignmentDetector()
eye_landmarks = idet.detect_iris(im)

5. Face verification

models.verifier.face_verifier.FaceVerifier(extractor="facenet", classes=512)

Argument

  • extractor: A string, one of facenet, insightface, ir50_ms1m, or ir50_asia.
  • classes: An integer. Dimension of output embeddings.

Example

from models.verifier import face_verifier

im1 = cv2.imread(PATH_TO_IMAGE1)[..., ::-1]
im2 = cv2.imread(PATH_TO_IMAGE2)[..., ::-1]
fv = face_verifier.FaceVerifier(extractor="facenet")
# fv.set_detector(fd) # fd = face_detector.FaceAlignmentDetector()
result, distance = fv.verify(im1, im2, threshold=0.5, with_detection=False, with_alignment=False, return_distance=True)

6. Gender and age estimation

models.estimator.gender_age_estimator.GenderAgeEstimator(model_type="insightface")

Arguments

  • model_type: A string, only insightface is supported now.

Example

from models.estimator import gender_age_estimator

im = cv2.imread(PATH_TO_IMAGE)[..., ::-1]
gae = gender_age_estimator.GenderAgeEstimator()
gae.set_detector(fd) # fd = face_detector.FaceAlignmentDetector()
gender, age = gae.predict_gender_age(im, with_detection=True)

Ported model weights

Known issues

It works fine on Colab at this point (2019/06/11) but for certain Keras/TensorFlow version, it throws errors loading 2DFAN-1_keras.h5 or 2DFAN-2_keras.h5.

Requirements

  • Keras 2.2.4
  • TensorFlow 1.12.0 or 1.13.1

Acknowledgments

We learnt a lot from 1adrianb/face-alignment, zllrunning/face-parsing.PyTorch, swook/GazeML, deepinsight/insightface, davidsandberg/facenet, and ZhaoJ9014/face.evoLVe.PyTorch.