yolov7-face

Method	Test Size	Easy	Medium	Hard	Link
yolov7-tiny-leak	640	93.2	91.3	83.0	google
yolov5s	640	94.0	92.1	84.5	google
yolov7-lite-e	640	91.1	87.5	70.9	google

Demo

Download the model place into weights folder, then run:

python detect.py --weights weights/yolov7-tiny-face.pt --view-img

Deploy

事实上这个仓库主要是验证wnnx在复杂模型下的推理能力，当然这当中存在着非常多的坑，最终，我们可以做到NCNN里面做不到的东西，后处理几乎全部融合在模型里面，外部只需要做一个score判断即可。

wnnx的推理结果：

第一张图为yolov7-tiny，后面两张图为 yolov7-lite-e，单线程在CPU下跑的realtime，多线程更快，fp16更快，精度还可以。

Export

export to onnx and torchscript:

python export.py --weights weights/yolov7-tiny-face.pt

Dataset

WiderFace

yolov7-face-label

Test

Demo

• ncnn_Android_face

References

• https://github.com/deepcam-cn/yolov5-face

• https://github.com/WongKinYiu/yolov7

• https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose

• https://github.com/ppogg/YOLOv5-Lite