PFLD - Ultralight - 400FPS(CPU)
!!2022年结合MobileOne的性能更加优异的人脸关键点检测算法已开源->PFLD_GhostOne
实现功能
- PFLD的训练/测试/评估/ncnn C++推理
- PFLD_Ulitralight的训练/测试/评估/ncnn C++推理
- 人脸98个关键点检测
- 支持onnx导出
基于PFLD进行优化的超轻量级人脸关键点检测器
提供了一系列适合移动端部署的人脸关键点检测器： 对PFLD网络的基础结构进行优化，并调整网络结构，使其更适合边缘计算，而且在绝大部分情况下精度均好于原始PFLD网络，在CPU下运行最快可达到400fps。
运行环境
- Ubuntu 16.04
- Python 3.7
- Pytorch 1.4
- CUDA 10.1
如果想省去环境搭建工作，可以使用以下Docker环境: Docker_PFLD
精度
WFLW测试结果
| 输入大小为112x112 | Model | Width | NME | Inference Time(NCNN 1xCore) | Model Size |
|---|---|---|---|---|---|
| PFLD | 0.25 1 |
0.06159 0.05837 |
5.5ms 39.7ms |
0.45M 5M |
|
| PFLD_Ultralight | 0.25 1 |
0.06101 (↓0.94%) 0.05749 (↓1.51%) |
3.6ms (↓34.5%) 18.6ms (↓53.1%) |
0.41M (↓8.89%) 3.40M (↓32.0%) |
|
| PFLD_Ultralight_Slim | 0.25 1 |
0.06258 (↑1.61%) 0.05627 (↓3.60%) |
3.3ms (↓40.0%) 16.0ms (↓59.7%) |
0.38M (↓15.6%) 3.10M (↓38.0%) |
| 输入大小为96x96 | Model | Width | NME | Inference Time(NCNN 1xCore) | Model Size |
|---|---|---|---|---|---|
| PFLD | 0.25 1 |
0.06136 0.05818 |
4.1ms 29.2ms |
0.43M 4.8M |
|
| PFLD_Ultralight | 0.25 1 |
0.06321 (↑3.0%) 0.05745 (↓1.25%) |
2.7ms (↓34.1%) 14.0ms (↓52.1%) |
0.39M (↓9.30%) 3.20M (↓33.3%) |
|
| PFLD_Ultralight_Slim | 0.25 1 |
0.06402 (↑4.33%) 0.05683 (↓2.32%) |
2.5ms (↓40.0%) 12.0ms (↓59.7%) |
0.37M (↓14.0%) 2.90M (↓39.6%) |
Installation
Clone and install:
- git clone https://github.com/AnthonyF333/PFLD_UltraLight.git
- cd ./PFLD_UltraLight
- pip install -r requirement.txt
- Pytorch version 1.4.0 are needed.
- Codes are based on Python 3.7
Data:
- Download WFLW dataset:
- Move the WFLW.zip to ./data/ directory and unzip the WFLW.zip
- Run SetPreparation.py to generate the training and test data.
- By default, it repeats 10 times for every image for augmentation, and save results in ./data/test_data/ and ./data/train_data/ directory.
Training
Before training, check or modify network configuration (e.g. batch_size, epoch and steps etc..) in config.py.
- MODEL_TYPE: you can choose PFLD, PFLD_Ultralight or PFLD_Ultralight_Slim for different network.
- WIDTH_FACTOR: you can choose 1 for original network or 0.25 for narrower network.
- TRAIN_DATA_PATH: the path of training data, by default it is ./data/train_data/list.txt which is generate by SetPreparation.py.
- VAL_DATA_PATH: the path of validation data, by default it is ./data/test_data/list.txt which is generate by SetPreparation.py.
After modify the configuration, run train.py to start training.
Test
Before test, modify the configuration in test.py, include the model_path, test_dataset, model_type, width_factor etc. Then run test.py to align the images in the test dataset, and save results in ./test_result directory.
C++ inference ncnn
1) Generate the onnx file: Modify the model_type, width_factor, model_path in pytorch2onnx.py, and then run pytorch2onnx.py to generate the xxx.onnx file. 2) Simplify the onnx file: First run "pip3 install onnx-simplifier" to install the tool, then run "python3 -m onnxsim xxx.onnx xxx_sim.onnx" to simpify the onnx file. 3) Transform onnx to ncnn format: Run "./onnx2ncnn xxx_sim.onnx xxx_sim.param xxx_sim.bin" to generate the ncnn files.