CenterFace Documentation

CenterFace is an anchor-free face detection and alignment algorithm designed for practical use on edge devices. It treats faces as points rather than using traditional anchor boxes.

There are two primary model versions:

• CenterFace: A model with a size of 7.3M.
• CenterFace-small: A lightweight version with a size of 2.3M that maintains performance comparable to the standard CenterFace model.

CenterFace and CenterFace-small demonstrate high accuracy on the WIDER FACE dataset and low inference latency on NVIDIA hardware.

Accuracy (WIDER FACE val set):

• CenterFace: Easy (0.935), Medium (0.924), Hard (0.875)
• CenterFace-small: Easy (0.931), Medium (0.924), Hard (0.870)

Inference Latency (NVIDIA RTX 2080TI):

• CenterFace (640×480): 5.5ms
• CenterFace-small (640×480): 4.4ms
• CenterFace (1920×1080): 8.7ms
• CenterFace-small (1920×1080): 7.3ms

Note: For single-scale inference (SIO) on the val set, CenterFace achieves: 92.2% (Easy), 91.1% (Medium), and 78.2% (Hard).

To run the provided demonstration script, navigate to the prj-tensorrt directory and execute demo.py using Python. For more detailed instructions, refer to the external tutorials.

cd CenterFace/prj-tensorrt
python demo.py

To use the Python version of CenterFace with TensorRT acceleration, ensure your environment meets the following requirements:

• TensorRT: version 6.0.1 or higher
• Python: version 3.7
• OpenCV: installed and available in your environment

To build and run the C++ version of CenterFace with NCNN, you need the following dependencies installed on your system:

• ncnn: The core inference framework.
• opencv (optional): Required if you wish to use OpenCV-related features or if the specific build configuration depends on it.

Follow these steps to build the project using CMake and run the demo executable:

1. Navigate to the project directory: cd CenterFace/prj-mnn
2. Create a build directory and compile: mkdir build && cd build && cmake .. && make -j3
3. Run the demo by providing the path to the MNN models and the target image path: ./demo <model_directory> <image_path>

cd CenterFace/prj-mnn
mkdir build && cd build && cmake .. && make -j3
./demo ../../models/mnn  your_image_path

Before compiling, you must specify the path to your MNN installation in the CMakeLists.txt file. Locate line 4 and update the DIR variable to point to your MNN directory.

set (DIR /home/mirror/workspace/mnn)

Follow these steps to build the project using CMake and run the provided demo executable. Ensure you are in the prj-ncnn directory before starting.

cd CenterFace/prj-ncnn
mkdir build && cd build && cmake && make
./demo ../../models/ncnn  your_image_path

To use the C++ version of CenterFace with MNN, you must have the following dependencies installed on your system:

• mnn
• opencv (version 4.1.1 or compatible)

To build the project and run the demonstration, follow these steps from the prj-opencv-cpp directory:

1. Create a build directory and compile using CMake and Make.
2. Run the ./demo executable, providing the path to your ONNX models and the path to the image you wish to process.

cd CenterFace/prj-opencv-cpp
mkdir build && cd build && cmake && make
./demo ../../models/onnx  your_image_path