retinaface_onnx

This repo is based on https://github.com/biubug6/Pytorch_Retinaface which updated with the ability of converting model backbone into onnx model, allow using cpu for inferencing in realtime (TLDR: It's fast when running on cpu)

Setup

1. Using conda to create a virtual environment

conda create --name py36 python=3.6
conda activate py36

2. install needed python lib

pip install -r requirements.txt

3. Download weight file here and put it in

/retinaface_onnx/module/face_detector/retinaface/weights/

4. Run scripts

• Change setting in file config.py according to your need
• Script running Realtime detect face:

  python infer.py

• Script convert pytorch model into onnx model

  python convert_onnx.py

Performence

On my laptop (CPU: AMD Ryzen 5 3500U, 12Gb Ram), the smallest backbone (MobileNet) archive performance of 30ms/image

Reference:

• https://github.com/biubug6/Pytorch_Retinaface
• https://pytorch.org/docs/stable/onnx.html
• onnx/onnx#2939
• https://github.com/pytorch/pytorch/blob/326d777e5384a621306330b5af0f2857843fe544/test/onnx/test_operators.py#L277