Retinaface Face Detection (Pytorch)

Authors: wxywb

Overview

This opertator detects faces in the images by using RetinaFace Detector[1]. It will returns the locations, five keypoints and the cropped face images from origin images. This repo is a adopataion from [2].

Interface

__call__(self, image: 'towhee.types.Image')

Args:

• image:
  • the image to detect faces.
  • supported types: towhee.types.Image

Returns:

The Operator returns a tupe Tuple[('boxes', numpy.ndarray), ('keypoints', numpy.ndarray), ('cropped_imgs', numpy.ndarray)])] containing following fields:

• boxes:
  • boxes of human faces.
  • data type: numpy.ndarray
  • shape: (num_faces, 4)
• keypoints:
  • keypoints of human faces.
  • data type: numpy.ndarray
  • shape: (10)
• cropped_imgs:
  • cropped face images.
  • data type: numpy.ndarray
  • shape: (h, w, 3)

Requirements

You can get the required python package by requirements.txt.

How it works

The towhee/retinaface-face-detection Operators implents the function of face detection. The example pipeline can be found in face-embedding-retinaface-inceptionresnetv1

Reference

[1]. https://arxiv.org/abs/1905.00641
[2]. https://github.com/biubug6/Pytorch_Retinaface