linzai

不用改

人脸大致为正方形

1.vino用的不多，不太清楚。 2.可能是由于模型比较小，而且主干是mb的,单位读取内存数据内的计算量不大，不能完全发挥gpu 的优势，所以和cpu差不多。

You can choose the face with the highest score.

You can use the [c++ version of MNN inference code](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/MNN) provided in my repo, which runs on 4 cores by default.

wins下训练问题可能的确比较多，我在linux下训练的。建议还是linux下训练，win下做推理问题不大。

There are currently no five landmarks now ，but a version containing it will be available soon.

I have been busy recently, and the model with 5 points landmarks that I have trained is not good enough, and I still need to adjust when I have time,...

I haven't researched onnx.js yet, you can try these solutions.

暂时还没有.