Face Detection Benchmark

Notes

Find all the models and datasets here https://drive.google.com/drive/folders/1C5OmbEOIcEPVOVLLF2PnyUYg7sp2Fx3O?usp=sharing

Version thats works

• Python: 3.12
• CUDA: 12.1
• cuDNN: 9.2

Use the steps below to run the WIDER Face detection benchmark with the CUDA benchmark script.

0. Create a Python 3.12 virtual environment and verify the system

From the repository root, run:

py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
python --version

Expected:

• Python 3.12.x

1. Download the WIDER Face dataset and evaluation code

• Download the WIDER Face dataset and the evaluation toolkit.
• Extract or copy them into datasets/widerface/.
• If you want to share the files from Google Drive, add your link here in the README.

The benchmark expects this structure:

datasets/widerface/
	evaluation.py
	setup.py
	WIDER_val/
	wider_face_eval_tools/

1.1 Build the WIDER Face evaluation extension

From the repository root, run:

cd datasets/widerface
python setup.py build_ext --inplace
cd ../..

2. Download the candidate face detection models

• Download the candidate ONNX models.
• Put them in models/face_detection/.
• If you want to share the model files from Google Drive, add your link here in the README.

The current CUDA benchmark script expects these files:

models/face_detection/scrfd10gkps.onnx
models/face_detection/face_detection_yunet_2023mar_raven.onnx

3. Install Python dependencies

From the repository root, run:

pip install -r requirements.txt

If you are using a virtual environment, activate it first.

3.1 Verify CUDA and ONNX Runtime

Run a quick runtime check:

python -c "import torch, onnxruntime as ort; print('Torch CUDA:', torch.cuda.is_available()); print('ORT providers:', ort.get_available_providers())"

Expected:

• Torch CUDA: True (if CUDA is set up correctly)
• ORT providers includes CUDAExecutionProvider

4. Run the CUDA benchmark

From the repository root, run:

python src/benchmarks/benchmark_detector_cuda.py

The script will prompt you to select a CUDA device if one is available.

Outputs

The benchmark writes logs and evaluation outputs under src/benchmarks/results/.