SavyaSanchi/opencv_contribution/yolov4 - Storage Bucket

YOLOv4, YOLOv4-tiny and YOLOv4x-mish ONNX Conversion

Prerequisites

1. Model files (cfg + weights)

The Darknet .cfg files are already provided in opencv_extra/testdata/dnn (yolov4.cfg, yolov4-tiny-2020-12.cfg, yolov4x-mish.cfg).

Download the matching .weights files 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 YOLOv4 YOLOv4-tiny-2020-12 YOLOv4x-mish

2. Python environment for pytorch-YOLOv4

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

Supported Python versions: 3.7 – 3.10.

conda create -n <env_name> python=<3.7-3.10> -y
conda activate <env_name>
pip install "torch<2.4" "torchvision<0.19" "numpy<2" onnx onnxruntime onnxscript

---

Conversion of YOLOv4 to ONNX

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

# Convert (dynamic batch, batch_size=0)
python -c "from tool.darknet2onnx import transform_to_onnx; transform_to_onnx('yolov4.cfg', 'yolov4.weights', 0)"

---

Conversion of YOLOv4-tiny to ONNX

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

# Convert YOLOv4-tiny (dynamic batch, batch_size=0)
python -c "from tool.darknet2onnx import transform_to_onnx; transform_to_onnx('yolov4-tiny-2020-12.cfg', 'yolov4-tiny.weights', 0)"

---

Conversion of YOLOv4x-mish to ONNX

Why it differs from YOLOv4

The yolov4x-mish.cfg uses new_coords=1 — a Scaled-YOLOv4 optimization. The network is trained to output values directly in the [0, 1] range (no sigmoid needed) and uses a squared formula (t_w * 2)² * anchor for width/height instead of exp(t_w) * anchor. This is more numerically stable for large models.

The Issue

The pytorch-YOLOv4 converter was written for the original YOLOv4 (new_coords=0) and always applies sigmoid + exp. With new_coords=1 weights, sigmoid gets applied on top of already-activated values, squishing confidences from ~0.93 down to ~0.36. As a result, the model produces garbage detections.

The Fix (Modified scripts provided)

To resolve this, modified versions of darknet2pytorch.py and yolo_layer.py are provided in this repository. These scripts include the following patches:

• darknet2pytorch.py: Properly reads the new_coords flag from the .cfg.
• yolo_layer.py: Skips the redundant sigmoid activation for xy/obj/cls and implements the squared wh formula when new_coords=1 is detected.

Conversion Steps

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

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

# Convert YOLOv4x-mish (dynamic batch, batch_size=0)
python -c "from tool.darknet2onnx import transform_to_onnx; transform_to_onnx('yolov4x-mish.cfg', 'yolov4x-mish.weights', 0)"

---

Usage

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

# YOLOv4 (input size: 608x608)
python demo.py --model yolov4.onnx --image example_outputs/input.jpg --output example_outputs/yolov4_output.jpg

# YOLOv4-tiny (input size: 416x416)
python demo.py --model yolov4-tiny.onnx --image example_outputs/input.jpg --output example_outputs/yolov4-tiny_output.jpg

# YOLOv4x-mish (input size: 640x640)
python demo.py --model yolov4x-mish.onnx --image example_outputs/input.jpg --output example_outputs/yolov4x-mish_output.jpg

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

---

License

See License.txt — This conversion tool is based on pytorch-YOLOv4 (Apache-2.0). Original YOLOv4 model weights and configuration are released by Alexey Bochkovskiy (AlexeyAB/darknet).