YOLOv3 ONNX Conversion

Prerequisites

1. Model files (cfg + weights)

The Darknet .cfg file is already provided in opencv_extra/testdata/dnn (yolov3.cfg).

Download the .weights file using the OpenCV test data download script:

git clone https://github.com/opencv/opencv_extra.git
cd opencv_extra/testdata/dnn
python download_models.py YOLOv3

2. Python environment for `pytorch-YOLOv4`

The conversion uses pytorch-YOLOv4. Create a Python environment with the required dependencies:

Supported Python versions: 3.8 – 3.10.

conda create -n <env_name> python=<3.8-3.10> -y
conda activate <env_name>
pip install "torch<2.4" "torchvision<0.19" "numpy<2" onnx onnxruntime "onnxscript==0.1.0"

Conversion of YOLOv3 to ONNX

Why it requires a patch

The original YOLOv3 .cfg does not contain the scale_x_y field (introduced in YOLOv4). The pytorch-YOLOv4 converter requires this field unconditionally, causing a KeyError when converting YOLOv3. The fix is to default scale_x_y to 1.0 when the field is absent.

The Fix (modified script provided)

A patched version of darknet2pytorch.py is provided in this repository. It adds a default value for scale_x_y when the field is absent:

# changed line in tool/darknet2pytorch.py
yolo_layer.scale_x_y = float(block.get('scale_x_y', 1.0))

Conversion Steps

git clone https://github.com/Tianxiaomo/pytorch-YOLOv4.git
cd pytorch-YOLOv4

# [!] Replace tool/darknet2pytorch.py with the patched version from this repository
#     before running the conversion.

# Convert YOLOv3 (dynamic batch, batch_size=0)
python -c "from tool.darknet2onnx import transform_to_onnx; transform_to_onnx('yolov3.cfg', 'yolov3.weights', 0)"

The output file will be named yolov4_-1_3_416_416_dynamic.onnx (the script uses yolov4 as the default prefix regardless of input model).

Usage

A demo script is provided to run inference using OpenCV DNN:

python demo.py --model yolov3.onnx \
               --image example_outputs/input.jpg \
               --output example_outputs/yolov3_output.jpg

The demo prints the detected COCO classes, confidence scores, and bounding boxes, and saves an annotated output image.

License

See LICENSE — This conversion tool is based on pytorch-YOLOv4 (Apache-2.0). Original YOLOv3 model weights and configuration are released by Joseph Redmon (pjreddie/darknet).