English | 简体中文
- SCRFD
- （1）The *.pt provided by the Official Library can be deployed after the Export ONNX Model to complete the deployment；
- （2）As for SCRFD model trained on customized data, please follow Export ONNX Model to complete the deployment.
# Download scrfd model files
e.g. download from https://onedrive.live.com/?authkey=%21ABbFJx2JMhNjhNA&id=4A83B6B633B029CC%215542&cid=4A83B6B633B029CC
# Install the official library to configure the environment. This version should be exported in the following environment:
- Configure the environment manually
torch==1.8.0
mmcv==1.3.5
mmdet==2.7.0
- Configure via docker
docker pull qyjdefdocker/onnx-scrfd-converter:v0.3
# Export files in onnx format
- Manual generation
python tools/scrfd2onnx.py configs/scrfd/scrfd_500m.py weights/scrfd_500m.pth --shape 640 --input-img face-xxx.jpg
- docker
onnx files are in docker's onnx directory
For developers' testing, models exported by SCRFD are provided below. Developers can download and use them directly. (The accuracy of the models in the table is sourced from the official library)
| Model | Size | Accuracy |
|---|---|---|
| SCRFD-500M-kps-160 | 2.5MB | - |
| SCRFD-500M-160 | 2.2MB | - |
| SCRFD-500M-kps-320 | 2.5MB | - |
| SCRFD-500M-320 | 2.2MB | - |
| SCRFD-500M-kps-640 | 2.5MB | 90.97% |
| SCRFD-500M-640 | 2.2MB | 90.57% |
| SCRFD-1G-160 | 2.5MB | - |
| SCRFD-1G-320 | 2.5MB | - |
| SCRFD-1G-640 | 2.5MB | 92.38% |
| SCRFD-2.5G-kps-160 | 3.2MB | - |
| SCRFD-2.5G-160 | 2.6MB | - |
| SCRFD-2.5G-kps-320 | 3.2MB | - |
| SCRFD-2.5G-320 | 2.6MB | - |
| SCRFD-2.5G-kps-640 | 3.2MB | 93.8% |
| SCRFD-2.5G-640 | 2.6MB | 93.78% |
| SCRFD-10G-kps-320 | 17MB | - |
| SCRFD-10G-320 | 15MB | - |
| SCRFD-10G-kps-640 | 17MB | 95.4% |
| SCRFD-10G-640 | 15MB | 95.16% |
| SCRFD-10G-kps-1280 | 17MB | - |
| SCRFD-10G-1280 | 15MB | - |
- This tutorial and related code are written based on SCRFD CommitID:17cdeab