【git-hub项目：YOLOs-CPP】本地实现03：跑自己的实例分割模型

本节博客，我们继续讲解，如何在cpu+windows上，跑通自己的实例分割模型。

完整项目一键下载【⬇️⬇️⬇️】：

【GitHub跑通项目：YOLOs-CPP】+【计算机视觉：完整项目】+【YOLOv5-v11模型】+【Cpp+ONNX+cpu部署】+【目标检测+实例分割】资源-CSDN文库

目录

模型

类别名称

量化

导出模型

拉取最新代码

进入官网ultralytics

---

模型

该项目包括存储在 models 和 quantized_models 目录中的各种预训练标准 YOLO 模型：

模型类型模型名称标准模型yolo5-n6.onnxyolo7-tiny.onnxyolo8n.onnxyolo8n-seg.onnxyolo10n.onnxyolo11n.onnxyolo11n-seg.onnx量化模型yolo5-n6_uint8.onnxyolo7-tiny-uint8.onnxyolo8n_uint8.onnxyolo8n-seg_uint8.onnxyolo10n_uint8.onnxyolo11n_uint8.onnxyolo11n-seg_uint8.onnx

你也可以使用带有自定义类别的自定义 YOLO 版本！

比如，我们用ultralytics训练好的实例分割模型best.pt，我们再用原生的ultralytics转onnx方式，转为best.onnx。如下：

类别名称

coco.names：包含模型所使用的类别标签列表。

这里，我们也要修改coco.names，改为自己的标签！

量化

quantized_models 目录中包含针对低精度推理优化的 YOLO 模型的量化版本。此外，quantized_models/yolos_quantization.py 脚本可用于对自定义 YOLO 模型进行自定义量化。

注意：量化模型具有模型体积更小的优势，并且在准确性略有下降的情况下，可能实现更快的推理速度。

我们先不量化试一下：

修改源码：

然后，点击重新生成！

并且，运行exe，但是，你会发现，没有出结果。

怎么办呢？我们查看项目源码，发现，后来支持了seg：

但是，源码还是det【目标检测】。

因此，我们将源码从原始代码：

/** * @file image_inference.cpp * @brief Object detection in a static image using YOLO models (v5, v7, v8, v10). * * This file implements an object detection application that utilizes YOLO * (You Only Look Once) models, specifically versions 5, 7, 8, and 10. * The application loads a specified image, processes it to detect objects, * and displays the results with bounding boxes around detected objects. * * The application supports the following functionality: * - Loading a specified image from disk. * - Initializing the YOLO detector with the desired model and labels. * - Detecting objects within the image. * - Drawing bounding boxes around detected objects and displaying the result. * * Configuration parameters can be adjusted to suit specific requirements: * - `isGPU`: Set to true to enable GPU processing for improved performance; * set to false for CPU processing. * - `labelsPath`: Path to the class labels file (e.g., COCO dataset). * - `imagePath`: Path to the image file to be processed (e.g., dogs.jpg). * - `modelPath`: Path to the desired YOLO model file (e.g., ONNX format). * * The application can be extended to use different YOLO versions by modifying * the model path and the corresponding detector class. * * Usage Instructions: * 1. Compile the application with the necessary OpenCV and YOLO dependencies. * 2. Ensure that the specified image and model files are present in the * provided paths. * 3. Run the executable to initiate the object detection process. * * @note The code includes commented-out sections to demonstrate how to switch * between different YOLO models and image inputs. * * Author: Abdalrahman M. Amer, .linkedin /in/abdalrahman-m-amer * Date: 29.09.2024 */ // Include necessary headers #include <opencv2/highgui/highgui.hpp> #include <iostream> #include <string> // #include "YOLO5.hpp" // Uncomment for YOLOv5 // #include "YOLO7.hpp" // Uncomment for YOLOv7 // #include "YOLO8.hpp" // Uncomment for YOLOv8 // #include "YOLO10.hpp" // Uncomment for YOLOv10 #include "YOLO11.hpp" // Uncomment for YOLOv11 int main() { // Paths to the model, labels, and test image const std::string labelsPath = "../../models/coco.names"; const std::string imagePath = "../../data/dogs.jpg"; // Primary image path // Uncomment the desired image path for testing // const std::string imagePath = "../data/happy_dogs.jpg"; // Alternate image // const std::string imagePath = "../data/desk.jpg"; // Another alternate image // Model paths for different YOLO versions // Uncomment the desired model path for testing // const std::string modelPath = "../models/yolo5-n6.onnx"; // YOLOv5 // const std::string modelPath = "../models/yolo7-tiny.onnx"; // YOLOv7 // const std::string modelPath = "../models/yolo8n.onnx"; // YOLOv8 // const std::string modelPath = "../quantized_models/yolo10n_uint8.onnx"; // Quantized YOLOv10 const std::string modelPath = "../../models/yolo11n.onnx"; // YOLOv11 // Initialize the YOLO detector with the chosen model and labels bool isGPU = true; // Set to false for CPU processing // YOLO7Detector detector(modelPath, labelsPath, isGPU); // YOLO5Detector detector(modelPath, labelsPath, isGPU); // Uncomment for YOLOv5 // YOLO8Detector detector(modelPath, labelsPath, isGPU); // Uncomment for YOLOv8 // YOLO10Detector detector(modelPath, labelsPath, isGPU); // Uncomment for YOLOv10 YOLO11Detector detector(modelPath, labelsPath, isGPU); // Uncomment for YOLOv11 // Load an image cv::Mat image = cv::imread(imagePath); if (image.empty()) { std::cerr << "Error: Could not open or find the image!\n"; return -1; } // Detect objects in the image and measure execution time auto start = std::chrono::high_resolution_clock::now(); std::vector<Detection> results = detector.detect(image); auto duration = std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now() - start); std::cout << "Detection completed in: " << duration.count() << " ms" << std::endl; // Draw bounding boxes on the image //detector.drawBoundingBox(image, results); // simple bbox drawing detector.drawBoundingBoxMask(image, results); // Uncomment for mask drawing // Display the image cv::imshow("Detections", image); cv::waitKey(0); // Wait for a key press to close the window return 0; }

修改为如下：

/** * @file image_inference.cpp */ // Include necessary headers #include <opencv2/highgui/highgui.hpp> #include <iostream> #include <string> #include "YOLO11Seg.hpp" // 确保包含实例分割的头文件 int main() { // Paths to the model, labels, and test image const std::string labelsPath = "../../models/coco.names"; const std::string imagePath = "../../data/1.jpg"; // Primary image path // Model paths for different YOLO versions // Uncomment the desired model path for testing // const std::string modelPath = "../quantized_models/yolo10n_uint8.onnx"; // Quantized YOLOv10 const std::string modelPath = "../../models/best.onnx"; // // Initialize the YOLO detector with the chosen model and labels bool isGPU = true; // Set to false for CPU processing YOLOv11SegDetector detector(modelPath, labelsPath, isGPU); // Load an image cv::Mat image = cv::imread(imagePath); if (image.empty()) { std::cerr << "Error: Could not open or find the image!\n"; return -1; } // 执行实例分割 auto start = std::chrono::high_resolution_clock::now(); std::vector<Segmentation> results = detector.segment(image); auto duration = std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now() - start); std::cout << "Segmentation completed in: " << duration.count() << " ms" << std::endl; // 绘制结果 // 可以选择绘制带掩码的检测框或仅绘制掩码 //detector.drawSegmentations(image, results); // 仅绘制掩码 detector.drawSegmentationsAndBoxes(image, results); //drawSegmentationsAndBoxes // Display the image cv::imshow("Detections", image); cv::waitKey(0); // 按任意键关闭窗口 return 0; }

再次编译，可以看到，成功啦！

然后，我们运行生成的exe。

有结果，但是，没有出图，什么原因呢？

我怀疑可能是模型有问题，因此，先使用原始项目提供的实例分割模型，进行了测试。首先，修改代码如下：  

/** * @file image_inference.cpp */ // Include necessary headers #include <opencv2/highgui/highgui.hpp> #include <iostream> #include <string> #include "YOLO11Seg.hpp" // 确保包含实例分割的头文件 int main() { // Paths to the model, labels, and test image const std::string labelsPath = "../../models/coco.names"; const std::string imagePath = "../../data/dog.jpg"; // Primary image path // Model paths for different YOLO versions // Uncomment the desired model path for testing // const std::string modelPath = "../quantized_models/yolo10n_uint8.onnx"; // Quantized YOLOv10 const std::string modelPath = "../../models/yolo11n-seg.onnx"; // // Initialize the YOLO detector with the chosen model and labels bool isGPU = false; // Set to false for CPU processing YOLOv11SegDetector detector(modelPath, labelsPath, isGPU); // Load an image cv::Mat image = cv::imread(imagePath); if (image.empty()) { std::cerr << "Error: Could not open or find the image!\n"; return -1; } // 执行实例分割 auto start = std::chrono::high_resolution_clock::now(); std::vector<Segmentation> results = detector.segment(image); auto duration = std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now() - start); std::cout << "Segmentation completed in: " << duration.count() << " ms" << std::endl; // 打印推理结果 std::cout << "Number of detected objects: " << results.size() << std::endl; for (const auto& seg : results) { std::cout << "Class ID: " << seg.classId << ", Confidence: " << seg.conf << std::endl; std::cout << "Bounding Box: (" << seg.box.x << ", " << seg.box.y << ")" << " width: " << seg.box.width << " height: " << seg.box.height << std::endl; if (seg.mask.empty()) { std::cout << "Mask is empty!" << std::endl; } else { std::cout << "Mask size: " << seg.mask.size() << std::endl; } } // 绘制结果 // 可以选择绘制带掩码的检测框或仅绘制掩码 //detector.drawSegmentations(image, results); // 仅绘制掩码 detector.drawSegmentationsAndBoxes(image, results); //drawSegmentationsAndBoxes // 保存图像到文件 cv::imwrite("result.jpg", image); // Display the image cv::imshow("Detections", image); cv::waitKey(0); // 按任意键关闭窗口 return 0; }

还原了coco.names。

person bicycle car motorbike aeroplane bus train truck boat traffic light fire hydrant stop sign parking meter bench bird cat dog horse sheep cow elephant bear zebra giraffe backpack umbrella handbag tie suitcase frisbee skis snowboard sports ball kite baseball bat baseball glove skateboard surfboard tennis racket bottle wine glass cup fork knife spoon bowl banana apple sandwich orange broccoli carrot hot dog pizza donut cake chair sofa pottedplant bed diningtable toilet tvmonitor laptop mouse remote keyboard cell phone microwave oven toaster sink refrigerator book clock vase scissors teddy bear hair drier toothbrush

再次编译，并运行，就会发现，可以出结果：

因此，我们的目标就定位在了我们的模型上，我们需要重新导出一个新的onnx模型，才能跑通。

导出模型 参考链接：模型导出参考

这给了我灵感，我需要通过官方代码，进行导出模型。

因此，我依次进行了如下操作：

拉取最新代码 git clone github /ultralytics/ultralytics.git 进入官网ultralytics

复制代码，然后，配置环境【按照官方文档】，将训练好的best.pt放到适当位置，进行模型导出。

使用导出后的模型进行预测。就会得到如下结果：

可以看到，生成了最后的预测结果！

开心，终于跑通了！

具体的cpp代码如下：

/** * @file image_inference.cpp */ // Include necessary headers #include <opencv2/highgui/highgui.hpp> #include <iostream> #include <string> #include "YOLO11Seg.hpp" // 确保包含实例分割的头文件 int main() { // Paths to the model, labels, and test image const std::string labelsPath = "../../models/coco.names"; const std::string imagePath = "../../data/4.jpg"; // Primary image path // Model paths for different YOLO versions // Uncomment the desired model path for testing // const std::string modelPath = "../quantized_models/yolo10n_uint8.onnx"; // Quantized YOLOv10 const std::string modelPath = "../../models/best.onnx"; // // Initialize the YOLO detector with the chosen model and labels bool isGPU = false; // Set to false for CPU processing YOLOv11SegDetector detector(modelPath, labelsPath, isGPU); // Load an image cv::Mat image = cv::imread(imagePath); if (image.empty()) { std::cerr << "Error: Could not open or find the image!\n"; return -1; } // 执行实例分割 auto start = std::chrono::high_resolution_clock::now(); std::vector<Segmentation> results = detector.segment(image, 0.0f, 0.05f); auto duration = std::chrono::duration_cast<std::chrono::milliseconds>( std::chrono::high_resolution_clock::now() - start); std::cout << "Segmentation completed in: " << duration.count() << " ms" << std::endl; // 打印推理结果 std::cout << "Number of detected objects: " << results.size() << std::endl; for (const auto& seg : results) { std::cout << "Class ID: " << seg.classId << ", Confidence: " << seg.conf << std::endl; std::cout << "Bounding Box: (" << seg.box.x << ", " << seg.box.y << ")" << " width: " << seg.box.width << " height: " << seg.box.height << std::endl; if (seg.mask.empty()) { std::cout << "Mask is empty!" << std::endl; } else { std::cout << "Mask size: " << seg.mask.size() << std::endl; } } // 绘制结果 // 可以选择绘制带掩码的检测框或仅绘制掩码 //detector.drawSegmentations(image, results); // 仅绘制掩码 detector.drawSegmentationsAndBoxes(image, results); //drawSegmentationsAndBoxes // 保存图像到文件 cv::imwrite("result.jpg", image); // Display the image cv::imshow("Detections", image); cv::waitKey(0); // 按任意键关闭窗口 return 0; }