RetinaFace Detector

I just want to use RetinaFace conveniently. Most if not all of the codes in this repository are adapted from https://github.com/biubug6/Pytorch_Retinaface.

Changes

• remove the torch stuffs
• add onnx based inference
• add auto model downloading
• adapt the detect script for easier usage
• add bounding box squaring option

Contents

• Installation
• Inference
• References

Installation

pip install git+https://github.com/ndgnuh/retinaface_detector

Install ONNX runtime:

pip install onnxruntime-gpu
# or
pip install onnxruntime

Inference

Quick start:

from retinaface_detector import FaceDetector
import cv2

detector = FaceDetector()
cap = cv2.VideoCapture(0)
_, frame = cap.read()
boxes, scores, landmarks = detector(frame)

Full settings:

FaceDetector(
  model: str = 'mobilenet', # model name, either 'mobilenet' or 'resnet50', anything in the `retinaface_detector/configs.py`
  onnx_providers: Collection[str] = onnxruntime.get_available_providers(),
  config: dict = None, # default to getattr(configs, model)
  confidence_threshold: float = 0.02, # refer to https://github.com/biubug6/Pytorch_Retinaface/blob/master/detect.py
  top_k: int = 5000, # refer to https://github.com/biubug6/Pytorch_Retinaface/blob/master/detect.py
  keep_top_k: int = 750, # refer to https://github.com/biubug6/Pytorch_Retinaface/blob/master/detect.py
  nms_threshold: float = 0.4, # refer to https://github.com/biubug6/Pytorch_Retinaface/blob/master/detect.py
  score_threshold: float = 0.6, # refer to https://github.com/biubug6/Pytorch_Retinaface/blob/master/detect.py, it's visualize_threshold there
  square_box: Optional[str] = None) # does nothing by default, see below for details

Square box

This options modify the bounding box to a square:

x1 = max(cx - sz, 0)
x2 = cx + sz
y1 = max(cy - sz, 0)
y2 = cy + sz

where cx, cy are the center coordinates of the box, sz is calculated based on the square_box argument:

• "max": max(w, h) // 2
• "min": min(w, h) // 2
• "avg": (w + h) // 4

If square_box is None, do nothing.

References

• FaceBoxes
• Retinaface (mxnet)
• Retinaface (torch)

@inproceedings{deng2019retinaface,
title={RetinaFace: Single-stage Dense Face Localisation in the Wild},
author={Deng, Jiankang and Guo, Jia and Yuxiang, Zhou and Jinke Yu and Irene Kotsia and Zafeiriou, Stefanos},
booktitle={arxiv},
year={2019}