retinaface - Deep Learning-Based Facial Detection and Landmark Identification

Introduction to RetinaFace

RetinaFace stands out as a state-of-the-art deep learning-based facial detection tool crafted specifically for Python users. It is renowned for its remarkable performance, even amidst crowded environments, making it a valuable asset for developers and researchers alike.

Background

RetinaFace serves as the face detection module within the broader InsightFace project. Initially crafted with MXNet, it has seen a successful transformation and re-implementation using TensorFlow by Stanislas Bertrand, a notable contributor to its development. This project is a streamlined version of Bertrand's work, retaining the core model structure and pre-trained weights while being transformed to be pip-installable for ease of use.

Installation

Getting started with RetinaFace is a breeze. The simplest method is through PyPI, where users can conveniently download and install the library along with its prerequisites using:

$ pip install retina-face

Once installed, users can import the library and start utilizing its functionalities within their Python projects.

Key Features

Face Detection

RetinaFace provides an efficient face detection function. By feeding an image's exact path to the API, users receive the facial coordinates, essential landmarks (such as eyes, nose, and mouth), and a confidence score for each detected face. This setup is impressive for tasks requiring precision and reliability.

resp = RetinaFace.detect_faces("img1.jpg")

Example Output:

{
    "face_1": {
        "score": 0.9993440508842468,
        "facial_area": [155, 81, 434, 443],
        "landmarks": {
          "right_eye": [257.82974, 209.64787],
          "left_eye": [374.93427, 251.78687],
          "nose": [303.4773, 299.91144],
          "mouth_right": [228.37329, 338.73193],
          "mouth_left": [320.21982, 374.58798]
        }
  }
}

Face Alignment

RetinaFace proves essential in a modern face recognition pipeline, offering face alignment features that potentially increase recognition accuracy by nearly 1%. The library efficiently finds and uses eye coordinates for aligning faces, an integral step before other face recognition tasks.

import matplotlib.pyplot as plt
faces = RetinaFace.extract_faces(img_path="img.jpg", align=True)
for face in faces:
  plt.imshow(face)
  plt.show()

Face Recognition

While RetinaFace focuses on face detection, it pairs seamlessly with ArcFace, the face recognition module of InsightFace. Together, these tools form the basis of the DeepFace library, offering a comprehensive end-to-end face recognition solution.

from deepface import DeepFace
obj = DeepFace.verify("img1.jpg", "img2.jpg", model_name='ArcFace', detector_backend='retinaface')
print(obj["verified"])

ArcFace boasts an impressive accuracy rate of 99.40% on the LFW dataset, surpassing even human accuracy levels.

Contribution

The RetinaFace project welcomes contributions. Developers interested in contributing should run unit tests and linting checks locally using make test and make lint before submitting their pull requests. GitHub Actions will then handle test workflows automatically.

Support and Acknowledgements

Supporting RetinaFace can be done in several ways, including starring the repository or becoming a patron via platforms like Patreon or GitHub Sponsors. This project builds upon the work of InsightFace and is inspired heavily by Stanislas Bertrand's implementation and Ross Girshick's Fast R-CNN contributions.

Licensing

RetinaFace is distributed under the MIT License. Users should refer to the LICENSE file for detailed terms and conditions.

For researchers and developers looking to incorporate advanced facial detection into their Python applications, RetinaFace offers a robust, reliable, and easy-to-use solution, building upon well-researched foundations and community contributions.