This project supplies a better face landmark detector that is suitable for embedding devices: PFLD_GhostOne which is more suitable for edge computing. In most cases, its accuracy is better than the original PFLD model, and its speed is around 55% faster than the original PFLD model.
The proposed GhostOne module is show below:
More details about GhostOne can be found on this blog
- Support training, testing and ONNX inference of PFLD_GhostOne model
- Support detecting 98 facial landmarks
- Support exporting to ONNX format
- Ubuntu 18.04
- Python 3.9
- Pytorch 1.11.0
- CUDA 11.3
Model input size is 112x112
| Model | NME | OpenVino Latency(ms) | NCNN Latency(ms) | ONNX Model Size(MB) |
|---|---|---|---|---|
| PFLD | 0.05438 | 1.65(CPU) 2.78(GPU) | 5.4(CPU) 5.1(GPU) | 4.66 |
| PFLD-GhostNet | 0.05347 | 1.79(CPU) 2.55(GPU) | 2.9(CPU) 5.3(GPU) | 3.09 |
| PFLD-GhostNet-Slim | 0.05410 | 2.11(CPU) 2.54(GPU) | 2.7(CPU) 5.2(GPU) | 2.83 |
| PFLD-GhostOne | 0.05207 | 1.79(CPU) 2.18(GPU) | 2.4(CPU) 5.0(GPU) | 2.71 |
The latency is the average time of running 1000 times on 11th Gen Intel(R) Core(TM) i5-11500
| Model | Pretrained Model |
|---|---|
| PFLD | pfld_best.pth |
| PFLD-GhostNet | pfld_ghostnet_best.pth |
| PFLD-GhostNet-Slim | pfld_ghostnet_slim_best.pth |
| PFLD-GhostOne | pfld_ghostone_best.pth |
Clone and install:
- git clone https://github.com/AnthonyF333/PFLD_GhostOne.git
- cd ./PFLD_GhostOne
- pip install -r requirement.txt
- Pytorch version 1.11.0 are needed.
- Codes are based on Python 3.9
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 80 times for every image for augmentation, and save images in ./data/test_data_repeat80/ and ./data/train_data_repeat80/ directory.
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_GhostNet or PFLD_GhostNet_Slim or PFLD_GhostOne for different network.
- TRAIN_DATA_PATH: the path of training data, by default it is ./data/train_data_repeat80/list.txt which is generate by SetPreparation.py.
- VAL_DATA_PATH: the path of validation data, by default it is ./data/test_data_repeat80/list.txt which is generate by SetPreparation.py.
After modify the configuration, run train.py to start training.
Modify the model_type and model_path in pytorch2onnx.py, and then run pytorch2onnx.py to generate the xxx.onnx file. It will optimize and simplify the ONNX file.
Before test the model, modify the configuration in onnx_inference.py, include the pfld_onnx_model, test_folder, save_result_folder etc. Then run onnx_inference.py to detect and align the images in the test_folder, and save results in save_result_folder directory.