RetinaFace C++

This is the C++ implementation of the Retinaface PyTorch model inference.

The Inference Engine used is the ONNX Runtime Framework, and the model is obtained through the PyTorch ONNX export API. A pretrained model is provided in this path with dynamic input shape. In the current implementation, images with any dimension above 640px are resized to 640x640 letterbox padding, due to anchors decoding limitation.

Dependencies

• OpenCV 3 or higher

Usage

Bellow are the usage instructions

Unzip Pretrained Model

If there isn't an intent to provide a custom trained model, the first step consists in extracting the pretrained model. The following command called from the root path of this repo unzips the model.

unzip model/retinaface_dynamic.zip -d model

Compiling the main example

If a custom trained model is intended to be used, modify the main script in order to provide the correct path for model loading.

Use the following instructions to compile the main script:

mkdir build && cd build
cmake ..

Testing on an image

If the compilation succeded a binary file called RetinaCpp will be generated into the build directory, which can be used to detect faces in an image using the following structure:

./RetinaCpp PATH_TO_IMAGE

where PATH_TO_IMAGE should be replaced with a path to an image file.

Results

Bellow there is an example of a detection result provided by this current implementation.