掌握ONNXRuntime Yolov8-OBB，轻松实现旋转目标检测！

效果

    YOLOv8中OBB(Oriented Bounding Box)模型的引入标志着物体检测的重要一步，特别是对于有角度或旋转的物体，提高了准确性并减少了各种应用中的背景噪声，如航空影像和文本检测。

模型信息

Model Properties
-------------------------
date：2024-02-26T08:38:44.171849
description：Ultralytics YOLOv8s-obb model trained on runs/DOTAv1.0-ms.yaml
author：Ultralytics
task：obb
license：AGPL-3.0 https://ultralytics.com/license
version：8.1.18
stride：32
batch：1
imgsz：[640, 640]
names：{0: 'plane', 1: 'ship', 2: 'storage tank', 3: 'baseball diamond', 4: 'tennis court', 5: 'basketball court', 6: 'ground track field', 7: 'harbor', 8: 'bridge', 9: 'large vehicle', 10: 'small vehicle', 11: 'helicopter', 12: 'roundabout', 13: 'soccer ball field', 14: 'swimming pool'}
---------------------------------------------------------------

Inputs
-------------------------
name：images
tensor：Float[1, 3, 640, 640]
---------------------------------------------------------------

Outputs
-------------------------
name：output0
tensor：Float[1, 20, 8400]
---------------------------------------------------------------

项目

代码

using Microsoft.ML.OnnxRuntime;
using Microsoft.ML.OnnxRuntime.Tensors;
using OpenCvSharp;
using OpenCvSharp.Dnn;
using System;
using System.Collections.Generic;
using System.Drawing;
using System.IO;
using System.Linq;
using System.Windows.Forms;

namespace Onnx_Yolov8_Demo
{
    public partial class Form1 : Form
    {
        public Form1()
        {
            InitializeComponent();
        }

        string fileFilter = "*.*|*.bmp;*.jpg;*.jpeg;*.tiff;*.tiff;*.png";
        string image_path = "";
        string classer_path;
        DateTime dt1 = DateTime.Now;
        DateTime dt2 = DateTime.Now;
        string model_path;
        Mat image;
        Mat result_image;
        public string[] class_lables;
        SessionOptions options;
        InferenceSession onnx_session;
        Tensor<float> input_tensor;
        List<NamedOnnxValue> input_container;
        IDisposableReadOnlyCollection<DisposableNamedOnnxValue> result_infer;
        DisposableNamedOnnxValue[] results_onnxvalue;

        Tensor<float> result_tensors;

        private void button1_Click(object sender, EventArgs e)
        {
            OpenFileDialog ofd = new OpenFileDialog();
            ofd.Filter = fileFilter;
            if (ofd.ShowDialog() != DialogResult.OK) return;
            pictureBox1.Image = null;
            image_path = ofd.FileName;
            pictureBox1.Image = new Bitmap(image_path);
            textBox1.Text = "";
            image = new Mat(image_path);
            pictureBox2.Image = null;
        }

        private void button2_Click(object sender, EventArgs e)
        {
            if (image_path == "")
            {
                return;
            }

            button2.Enabled = false;
            pictureBox2.Image = null;
            textBox1.Text = "";
            Application.DoEvents();

            //图片缩放
            image = new Mat(image_path);
            int max_image_length = image.Cols > image.Rows ? image.Cols : image.Rows;
            Mat max_image = Mat.Zeros(new OpenCvSharp.Size(max_image_length, max_image_length), MatType.CV_8UC3);
            Rect roi = new Rect(0, 0, image.Cols, image.Rows);
            image.CopyTo(new Mat(max_image, roi));

            float[] result_array;
            float factor = (float)(max_image_length / 640.0);

            // 将图片转为RGB通道
            Mat image_rgb = new Mat();
            Cv2.CvtColor(max_image, image_rgb, ColorConversionCodes.BGR2RGB);
            Mat resize_image = new Mat();
            Cv2.Resize(image_rgb, resize_image, new OpenCvSharp.Size(640, 640));

            // 输入Tensor
            for (int y = 0; y < resize_image.Height; y++)
            {
                for (int x = 0; x < resize_image.Width; x++)
                {
                    input_tensor[0, 0, y, x] = resize_image.At<Vec3b>(y, x)[0] / 255f;
                    input_tensor[0, 1, y, x] = resize_image.At<Vec3b>(y, x)[1] / 255f;
                    input_tensor[0, 2, y, x] = resize_image.At<Vec3b>(y, x)[2] / 255f;
                }
            }

            //将 input_tensor 放入一个输入参数的容器，并指定名称
            input_container.Add(NamedOnnxValue.CreateFromTensor("images", input_tensor));

            dt1 = DateTime.Now;
            //运行 Inference 并获取结果
            result_infer = onnx_session.Run(input_container);
            dt2 = DateTime.Now;

            // 将输出结果转为DisposableNamedOnnxValue数组
            results_onnxvalue = result_infer.ToArray();

            // 读取第一个节点输出并转为Tensor数据
            result_tensors = results_onnxvalue[0].AsTensor<float>();

            result_array = result_tensors.ToArray();

            Mat result_data = new Mat(20, 8400, MatType.CV_32F, result_array);
            result_data = result_data.T();
            List<Rect2d> position_boxes = new List<Rect2d>();
            List<int> class_ids = new List<int>();
            List<float> confidences = new List<float>();
            List<float> rotations = new List<float>();
            // Preprocessing output results
            for (int i = 0; i < result_data.Rows; i++)
            {
                Mat classes_scores = new Mat(result_data, new Rect(4, i, 15, 1));
                OpenCvSharp.Point max_classId_point, min_classId_point;
                double max_score, min_score;
                // Obtain the maximum value and its position in a set of data
                Cv2.MinMaxLoc(classes_scores, out min_score, out max_score, out min_classId_point, out max_classId_point);
                // Confidence level between 0 ~ 1
                // Obtain identification box information
                if (max_score > 0.25)
                {
                    float cx = result_data.At<float>(i, 0);
                    float cy = result_data.At<float>(i, 1);
                    float ow = result_data.At<float>(i, 2);
                    float oh = result_data.At<float>(i, 3);
                    double x = (cx - 0.5 * ow) * factor;
                    double y = (cy - 0.5 * oh) * factor;
                    double width = ow * factor;
                    double height = oh * factor;
                    Rect2d box = new Rect2d();
                    box.X = x;
                    box.Y = y;
                    box.Width = width;
                    box.Height = height;
                    position_boxes.Add(box);
                    class_ids.Add(max_classId_point.X);
                    confidences.Add((float)max_score);
                    rotations.Add(result_data.At<float>(i, 19));
                }
            }
            // NMS 
            int[] indexes = new int[position_boxes.Count];
            CvDnn.NMSBoxes(position_boxes, confidences, 0.25f, 0.7f, out indexes);
            List<RotatedRect> rotated_rects = new List<RotatedRect>();
            for (int i = 0; i < indexes.Length; i++)
            {
                int index = indexes[i];
                float w = (float)position_boxes[index].Width;
                float h = (float)position_boxes[index].Height;
                float x = (float)position_boxes[index].X + w / 2;
                float y = (float)position_boxes[index].Y + h / 2;
                float r = rotations[index];
                float w_ = w > h ? w : h;
                float h_ = w > h ? h : w;
                r = (float)((w > h ? r : (float)(r + Math.PI / 2)) % Math.PI);
                RotatedRect rotate = new RotatedRect(new Point2f(x, y), new Size2f(w_, h_), (float)(r * 180.0 / Math.PI));
                rotated_rects.Add(rotate);
            }

            result_image = image.Clone();

            for (int i = 0; i < indexes.Length; i++)
            {
                int index = indexes[i];
                Point2f[] points = rotated_rects[i].Points();

                for (int j = 0; j < 4; j++)
                {
                    Cv2.Line(result_image, (OpenCvSharp.Point)points[j], (OpenCvSharp.Point)points[(j + 1) % 4], new Scalar(0, 255, 0), 2);
                }

                Cv2.PutText(result_image, class_lables[class_ids[index]] + "-" + confidences[index].ToString("0.00"),
                    (OpenCvSharp.Point)points[0], HersheyFonts.HersheySimplex, 0.8, new Scalar(0, 0, 255), 2);
            }

            pictureBox2.Image = new Bitmap(result_image.ToMemoryStream());
            textBox1.Text = "推理耗时:" + (dt2 - dt1).TotalMilliseconds + "ms";

            button2.Enabled = true;
        }

        private void Form1_Load(object sender, EventArgs e)
        {
            model_path = "model/yolov8s-obb.onnx";
            classer_path = "model/lable.txt";

            // 创建输出会话，用于输出模型读取信息
            options = new SessionOptions();
            options.LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_INFO;
            options.AppendExecutionProvider_CPU(0);// 设置为CPU上运行

            // 创建推理模型类，读取本地模型文件
            onnx_session = new InferenceSession(model_path, options);//model_path 为onnx模型文件的路径

            // 输入Tensor
            input_tensor = new DenseTensor<float>(new[] { 1, 3, 640, 640 });
            // 创建输入容器
            input_container = new List<NamedOnnxValue>();

            List<string> str = new List<string>();
            StreamReader sr = new StreamReader(classer_path);
            string line;
            while ((line = sr.ReadLine()) != null)
            {
                str.Add(line);
            }
            class_lables = str.ToArray();

            image_path = "test_img/1.png";
            pictureBox1.Image = new Bitmap(image_path);
            image = new Mat(image_path);
        }

    }
}