Fast_YOLOv8_CPP

Introduction

This repository contains the code for performing inference on the YOLOv8n object detection model in C++. The three approaches implemented are:

1. OpenCV DNN Module – The easiest to integrate, supports Intel OpenVINO backend for acceleration.
2. ONNX Runtime – More flexible across different platforms; allows selecting execution providers like OpenVINO or TensorRT based on hardware.
3. OpenVINO – Optimized for Intel hardware, delivering the best performance.

For comparison, I tested all three approaches using YOLOv8n object detection with both FP32 models and INT8 static quantized models.
OpenVINO proved to be the fastest on my Intel® Core™ i7-12650H processor.

Usage

Check out the subfolders for individual implementation instructions:

• opencv_dnn/ – OpenCV DNN Module.
• onnxruntime/ – ONNX Runtime with OpenVINO Execution Provider.
• openvino/ – Native OpenVINO Inference.

Result

Comparison of inference performance across different approaches:

Acknowledgment

• Ultralytics.
• ONNX Runtime Inference.