pfld_106_face_landmarks

106点人脸关键点检测的PFLD算法实现

• 😊ncnn cpp推理代码

• 转换后的ONNX模型

• 预训练权重

• 性能测试

• update GhostNet

• update MobileNetV3

  Backbone	param	MACC	nme	Link	ONNX
  MobileNetV2	1.26M	393M	4.96%	v2	v2.onnx
  MobileNetV3	1.44M	201.8M	4.40%	v3	v3.onnx
  MobileNetV3_Small	0.22M	13M	6.22%	lite	lite.onnx

测试电脑MacBook 2017 13-Inch CPU i5-3.1GHz (single core)

backbone	FPS(onnxruntime cpu)	Time(single face)
v2.onnx	60.9	16ms
V3.onnx	62.7	15.9ms
lite.onnx	255	3.9ms

• Requirements

    torch=1.2.0
    torchvision
    opencv-python
    tqdm
    onnxruntime==1.2.2
    numpy

• 数据集准备

  # 下载数据集到data/imgs下
  cd data
  python prepare.py

  # data 文件夹结构
  data/
    imgs/
    train_data/
      imgs/
      list.txt
    test_data/
      imgs/
      list.txt

• 训练

CUDA_VISIBLE_DEVICES=0 python train.py --backbone=v3
# 可选backbone为v2 v3 lite

• 结果 (MobileNetV2)

  

  

  

• Thanks

https://github.com/polarisZhao/PFLD-pytorch

https://github.com/microsoft/onnxruntime

https://github.com/kuan-wang/pytorch-mobilenet-v3