dbface-on-openvino
dbface-on-openvino copied to clipboard
yas-sim
Describes how to run DBFace, a real-time, single-shot face detection model on Intel OpenVINO
Overview
The objective of this project is to run the DBFace, a real-time, single-stage face detector on Intel(r) Distribution of OpenVINO(tm) Toolkit.
Original DBFace GitHub site:
https://github.com/dlunion/DBFace
Thie original developer used PyTorch to train the model. You need to convert the PyTorch model to OpenVINO IR model. You need to run 2 conversion steps to get the final IR model.
PyTorch (.pth) -> ONNX (.onnx) -> IR (.xml, .bin)
このプロジェクトの目的はリアルタイムでシングルステージの顔検出モデル、DBFaceをIntel(r) Distribution of OpenVINO(tm) Toolkit上で実行できるようにすることです。
Original DBFace GitHub site:
https://github.com/dlunion/DBFace
元の開発者はPyTorchを使ってモデルの学習を行っています。そのためPyTorchモデルからOpenVINO IRモデルへの変換を行わなくてはなりません。 IRモデルを得るためには２つの変換ステップを要します。
PyTorch (.pth) -> ONNX (.onnx) -> IR (.xml, .bin)
Detection result with a 4VGA (1280x960) input picture
1. Prerequisites
- Intel Distribution of OpenVINO toolkit 2020.2 (or newer)
- Model Downloader / Model Converter setup
Model Downloader guide
pip3 -r ${INTEL_OPENVINO_DIR}/deployment_tools/tools/model_downloader/requirements.in
pip3 -r ${INTEL_OPENVINO_DIR}/deployment_tools/tools/model_downloader/requirements-pytorch.in
If you fail to install PyTorch, go to PyTorch official web site and follow the QUICK START LOCALLY guide to install it. You need to have PyTorch >=1.4.
2. Download DBFace PyTorch model and weight
Download DBFace.py and dbface.pth from original developer's GitHub page.
3. Convert PyTorch model into ONNX model
Use pytorch_to_onnx.py in the model_downloader directory.
python3 ${INTEL_OPENVINO_DIR}/deployment_tools/tools/model_downloader/pytorch_to_onnx.py \
--model-name DBFace \
--weights dbface.pth \
--import-module DBFace \
--input-shape 1,3,320,320 \
--output-file dbface.onnx \
--input-names x \
--output-names sigmoid_hm,tlrb,landmark
Hint: You can get the input and output node names from the model source code (in this case, DBFace.py in the model directory)
def forward(self, x):
out = self.hs1(self.bn1(self.conv1(x)))
:
return sigmoid_hm, tlrb, landmark
4. Convert ONNX model into OpenVINO IR model
Use Model Optimizer (MO) to convert the ONNX model into IR model.
python3 ${INTEL_OPENVINO_DIR}/deployment_tools/model_optimizer/mo.py \
--input_model dbface.onnx \
--mean_values [180,154,150] \
--scale_values [73.7,70.0,70.9] \
--input_shape [1,3,960,1280] \
--output_dir dbface-4vga \
--data_type FP16
Hint: You can change the input shape with --input_shape option.
Hint: You can find the appropriate parameters (--mean_values and --scale_values) for input data preprocessing from the original source code (in this case, line 36-40 in main.py from the origianl GitHub site)
mean = [0.408, 0.447, 0.47]
std = [0.289, 0.274, 0.278]
image = common.pad(image)
image = ((image / 255.0 - mean) / std).astype(np.float32)
5. Run sample program
python3 dbface-infer.py -m model.xml -i input_image
A webCam #0 will be used when you specify 'cam' as the input file name.
output.jpg will be ganarated in the current directry. (non-webCam version only)
Command line example:
$ python3 dbface-infer.py -m dbface-4vga/dbface.xml -i image.jpg
$ python3 dbface-infer.py -m dbface-vga/dbface.xml -i cam
6. Test Environment
- Ubuntu 18.04 / Windows 10 1909
- OpenVINO 2020.3 LTS
yas-sim