OpenCV  4.9.0-dev
Open Source Computer Vision
No Matches

Prev Tutorial: OpenCV usage with OpenVINO
Next Tutorial: How to run deep networks in browser

Original author Alessandro de Oliveira Faria
Extended by Abduragim Shtanchaev
Compatibility OpenCV >= 4.9.0

Running pre-trained YOLO model in OpenCV

Deploying pre-trained models is a common task in machine learning, particularly when working with hardware that does not support certain frameworks like PyTorch. This guide provides a comprehensive overview of exporting pre-trained YOLO family models from PyTorch and deploying them using OpenCV's DNN framework. For demonstration purposes, we will focus on the YOLOX model, but the methodology applies to other supported models.

Currently, OpenCV supports the following YOLO models:

This support includes pre and post-processing routines specific to these models. While other older version of YOLO are also supported by OpenCV in Darknet format, they are out of the scope of this tutorial.

Assuming that we have successfully trained YOLOX model, the subsequent step involves exporting and running this model with OpenCV. There are several critical considerations to address before proceeding with this process. Let's delve into these aspects.

YOLO's Pre-proccessing & Output

Understanding the nature of inputs and outputs associated with YOLO family detectors is pivotal. These detectors, akin to most Deep Neural Networks (DNN), typically exhibit variation in input sizes contingent upon the model's scale.

Model Scale Input Size
Small Models 1 416x416
Midsize Models 2 640x640
Large Models 3 1280x1280

This table provides a quick reference to understand the different input dimensions commonly used in various YOLO models inputs. These are standard input shapes. Make sure you use input size that you trained model with, if it is differed from from the size mentioned in the table.

The next critical element in the process involves understanding the specifics of image pre-processing for YOLO detectors. While the fundamental pre-processing approach remains consistent across the YOLO family, there are subtle yet crucial differences that must be accounted for to avoid any degradation in performance. Key among these are the resize type and the padding value applied post-resize. For instance, the YOLOX model utilizes a LetterBox resize method and a padding value of 114.0. It is imperative to ensure that these parameters, along with the normalization constants, are appropriately matched to the model being exported.

Regarding the model's output, it typically takes the form of a tensor with dimensions [BxNxC+5] or [BxNxC+4], where 'B' represents the batch size, 'N' denotes the number of anchors, and 'C' signifies the number of classes (for instance, 80 classes if the model is trained on the COCO dataset). The additional 5 in the former tensor structure corresponds to the objectness score (obj), confidence score (conf), and the bounding box coordinates (cx, cy, w, h). Notably, the YOLOv8 model's output is shaped as [BxNxC+4], where there is no explicit objectness score, and the object score is directly inferred from the class score. For the YOLOX model, specifically, it is also necessary to incorporate anchor points to rescale predictions back to the image domain. This step will be integrated into the ONNX graph, a process that we will detail further in the subsequent sections.

PyTorch Model Export

Now that we know know the parameters of the pre-precessing we can go on and export the model from Pytorch to ONNX graph. Since in this tutorial we are using YOLOX as our sample model, lets use its export for demonstration purposes (the process is identical for the rest of the YOLO detectors). To exporting YOLOX we can just use export script. Particularly we need following commands:

git clone
wget # download pre-trained weights
python3 -m tools.export_onnx --output-name yolox_s.onnx -n yolox-s -c yolox_s.pth --decode_in_inference

NOTE: Here --decode_in_inference is to include anchor box creation in the ONNX graph itself. It sets this value to True, which subsequently includes anchor generation function.

Below we demonstrated the minimal version of the export script (which could be used for models other than YOLOX) in case it is needed. However, usually each YOLO repository has predefined export script.

import onnx
import torch
from onnxsim import simplify
# load the model state dict
ckpt = torch.load(ckpt_file, map_location="cpu")
# prepare dummy input
dummy_input = torch.randn(args.batch_size, 3, exp.test_size[0], exp.test_size[1])
#export the model
dynamic_axes={"input": {0: 'batch'},
"output": {0: 'batch'}})
# use onnx-simplifier to reduce reduent model.
onnx_model = onnx.load(args.output_name)
model_simp, check = simplify(onnx_model)
assert check, "Simplified ONNX model could not be validated", args.output_name)

Running Yolo ONNX detector with OpenCV Sample

Once we have our ONNX graph of the model, we just simply can run with OpenCV's sample. To that we need to make sure:

  1. OpenCV is build with -DBUILD_EXAMLES=ON flag.
  2. Navigate to the OpenCV's build directory
  3. Run the following command:
./bin/example_dnn_yolo_detector --input=<path_to_your_input_file> \
--classes=<path_to_class_names_file> \
--thr=<confidence_threshold> \
--nms=<non_maximum_suppression_threshold> \
--mean=<mean_normalization_value> \
--scale=<scale_factor> \
--yolo=<yolo_model_version> \
--padvalue=<padding_value> \
--paddingmode=<padding_mode> \
--backend=<computation_backend> \


  • –input: File path to your input image or video. If omitted, it will capture frames from a camera.
  • –classes: File path to a text file containing class names for object detection.
  • –thr: Confidence threshold for detection (e.g., 0.5).
  • –nms: Non-maximum suppression threshold (e.g., 0.4).
  • –mean: Mean normalization value (e.g., 0.0 for no mean normalization).
  • –scale: Scale factor for input normalization (e.g., 1.0).
  • –yolo: YOLO model version (e.g., YOLOv3, YOLOv4, etc.).
  • –padvalue: Padding value used in pre-processing (e.g., 114.0).
  • –paddingmode: Method for handling image resizing and padding. Options: 0 (resize without extra processing), 1 (crop after resize), 2 (resize with aspect ratio preservation).
  • –backend: Selection of computation backend (0 for automatic, 1 for Halide, 2 for OpenVINO, etc.).
  • –target: Selection of target computation device (0 for CPU, 1 for OpenCL, etc.).
  • –device: Camera device number (0 for default camera). If --input is not provided camera with index 0 will used by default.

Here mean, scale, padvalue, paddingmode should exactly match those that we discussed in pre-processing section in order for the model to match result in PyTorch

To demonstrate how to run OpenCV YOLO samples without your own pretrained model, follow these instructions:

  1. Ensure Python is installed on your platform.
  2. Confirm that OpenCV is built with the -DBUILD_EXAMPLES=ON flag.

Run the YOLOX detector(with default values):

git clone
cd opencv_extra/testdata/dnn
python yolox_s_inf_decoder
cd ..
cd <build directory of OpenCV>

This will execute the YOLOX detector with your camera. For YOLOv8 (for instance), follow these additional steps:

cd opencv_extra/testdata/dnn
python yolov8
cd ..
cd <build directory of OpenCV>
./bin/example_dnn_yolo_detector --model=onnx/models/yolov8n.onnx --yolo=yolov8 --mean=0.0 --scale=0.003921568627 --paddingmode=2 --padvalue=144.0 --thr=0.5 --nms=0.4 --rgb=0

Building a Custom Pipeline

Sometimes there is a need to make some custom adjustments in the inference pipeline. With OpenCV DNN module this is also quite easy to achieve. Below we will outline the sample implementation details:

  • Import required libraries
#include <opencv2/dnn.hpp>
#include <fstream>
#include <sstream>
#include "iostream"
#include "common.hpp"
  • Read ONNX graph and create neural network model:
Net net = readNet(weightPath);
int backend = parser.get<int>("backend");
  • Read image and pre-process it:
float paddingValue = parser.get<float>("padvalue");
bool swapRB = parser.get<bool>("rgb");
int inpWidth = parser.get<int>("width");
int inpHeight = parser.get<int>("height");
Scalar scale = parser.get<float>("scale");
Scalar mean = parser.get<Scalar>("mean");
ImagePaddingMode paddingMode = static_cast<ImagePaddingMode>(parser.get<int>("paddingmode"));
Size size(inpWidth, inpHeight);
Image2BlobParams imgParams(
// rescale boxes back to original image
Image2BlobParams paramNet;
paramNet.scalefactor = scale;
paramNet.size = size;
paramNet.mean = mean;
paramNet.swapRB = swapRB;
paramNet.paddingmode = paddingMode;
#define CV_32F
Definition interface.h:78
inp = blobFromImageWithParams(img, imgParams);
  • Inference:
std::vector<Mat> outs;
std::vector<int> keep_classIds;
std::vector<float> keep_confidences;
std::vector<Rect2d> keep_boxes;
std::vector<Rect> boxes;
net.forward(outs, net.getUnconnectedOutLayersNames());
  • Post-Processing

All post-processing steps are implemented in function yoloPostProcess. Please pay attention, that NMS step is not included into onnx graph. Sample uses OpenCV function for it.

outs, keep_classIds, keep_confidences, keep_boxes,
confThreshold, nmsThreshold,
  • Draw predicted boxes
for (auto box : keep_boxes)
boxes.push_back(Rect(cvFloor(box.x), cvFloor(box.y), cvFloor(box.width - box.x), cvFloor(box.height - box.y)));
paramNet.blobRectsToImageRects(boxes, boxes, img.size());
for (size_t idx = 0; idx < boxes.size(); ++idx)
Rect box = boxes[idx];
drawPrediction(keep_classIds[idx], keep_confidences[idx], box.x, box.y,
box.width + box.x, box.height + box.y, img);
const std::string kWinName = "Yolo Object Detector";
namedWindow(kWinName, WINDOW_NORMAL);
imshow(kWinName, img);
int cvFloor(double value)
Rounds floating-point number to the nearest integer not larger than the original.
Definition fast_math.hpp:231