C# OpenVinoSharp部署Yolo26模型进行推理

使用开源没多久的Yolo26模型进行推理测试，无需NMS操作

效果如下：

模型信息：

输出格式说明→ $1,300,6$ 6: x1 y1 x2 y2 score classeId(左上角右下角置信度类别ID)

代码如下：

using OpenCvSharp;

using OpenCvSharp.Extensions;

using OpenVinoSharp;

using OpenVinoSharp.Extensions.process;

using System;

using System.Collections.Generic;

using System.Data;

using System.IO;

using System.Linq;

using System.Windows.Forms;

namespace OPENVINOSHARP_NEW

{

public partial class Form1 : Form

{

public Form1()

{

InitializeComponent();

}

Core core;

CompiledModel compiledModel;

InferRequest inferRequest;

//分配随机颜色

private static readonly Scalar\[\] Colors = Enumerable.Repeat(false, 80).Select(x => Scalar.RandomColor()).ToArray();

//coco names

private static readonly string\[\] Labels = File.ReadAllLines("D:\\ultralytics-main\\classes.txt");

private void button1_Click(object sender, EventArgs e)

{

core = new Core();

compiledModel = core.compile_model("C:\\yolo26s.onnx", "CPU");

inferRequest = compiledModel.create_infer_request();

}

private void button2_Click(object sender, EventArgs e)

{

Mat image = new Mat("D:\\ultralytics-main\\bus.jpg");

Mat rgbMat = image.CvtColor(ColorConversionCodes.BGR2RGB);

rgbMat = OpenVinoSharp.Extensions.process.Resize.letterbox_img(rgbMat, 640, out var scales);

Mat normalizedMat = new Mat();

rgbMat.ConvertTo(normalizedMat, MatType.CV_32FC3, 1.0 / 255.0f); ;

float\[\] data = Permute.run(normalizedMat);

Tensor inputTensor = inferRequest.get_input_tensor();

inputTensor.set_data(data);

inferRequest.infer();

var outputTensor = inferRequest.get_output_tensor();

var output = outputTensor.get_data<float>((int)outputTensor.get_size()); // $1,300,6$ 6: x y w h score classe

var proposals = (int)outputTensor.get_shape() $1$ ;

var nums = (int)outputTensor.get_shape() $2$ ;

List<DetResult> results = new List<DetResult>();

for (int i = 0; i < proposals; i++)

{

float score = output $i \* nums + 4$ ;

if (score > 0.5)

{

int x1 = (int)(output $i \* nums + 0$ * scales);

int y1 = (int)(output $i \* nums + 1$ * scales);

int x2 = (int)(output $i \* nums + 2$ * scales);

int y2 = (int)(output $i \* nums + 3$ * scales);

Rect rect = new Rect(x1, y1, x2-x1, y2-y1);

int classId = (int)output $i \* nums + 5$ ;

DetResult detResult = new DetResult

{

box = rect,

score = score,

label = classId

};

results.Add(detResult);

}

// 绘制边界框

foreach (DetResult detResult in results)

{

Cv2.Rectangle(image, detResult.box, Colors $detResult.label$ , 2);

// 添加标签和置信度

string label = $"Class: {Labels $detResult.label$ } - {detResult.score:F2}";

Cv2.PutText(image, label,new Point(detResult.box.X, detResult.box.Y - 10),HersheyFonts.HersheySimplex, 0.5, new Scalar(255, 255, 0), 1);

}

pictureBox1.Image = image.ToBitmap();

rgbMat.Dispose();

normalizedMat.Dispose();

image.Dispose();

}

public class DetResult

{

public Rect box;

public float score;

public int label;

}