使用c++推理yolov8的图像分割模型

依赖库

opencv4.7及以上

导出onnx

yolo export model=./runs/segment/train3/weights/best.pt imgsz=640 format=onnx opset=12

cpp推理代码

#include<opencv.hpp>
#include<iostream>
#include<fstream>
#include<filesystem>
#include<random>

/*
yolov8图像分割推理
*/

std::vector<std::string> parse_classes_file(const char* name)
{
	std::vector<std::string> classes;

	std::ifstream file(name);
	if (!file.is_open()) 
	{
		std::cerr << "Error: fail to open classes file: " << name << std::endl;
		return classes;
	}

	std::string line;
	while (std::getline(file, line)) 
	{
		auto pos = line.find_first_of(" ");
		classes.emplace_back(line.substr(0, pos));
	}

	file.close();
	return classes;
}

cv::Mat modify_image_size(const cv::Mat& img)
{
	auto max = std::max(img.rows, img.cols);
	cv::Mat ret = cv::Mat::zeros(max, max, CV_8UC3);
	img.copyTo(ret(cv::Rect(0, 0, img.cols, img.rows)));

	return ret;
}

void get_masks(const cv::Mat& features, const cv::Mat& proto, const std::vector<int>& output1_sizes, const cv::Mat& frame, const cv::Rect box, cv::Mat& mk, const int input_size[2], const double mask_threshold)
{
	const cv::Size shape_src(frame.cols, frame.rows), // 原始图像大小
		shape_input(input_size[1], input_size[0]), // 640
		shape_mask(output1_sizes[3], output1_sizes[2]); //160
	
	std::cout << features.size << ", " << proto.size << std::endl;

	// 矩阵乘法
	cv::Mat res = (features * proto).t(); //1*25600 -> 25600*1

	res = res.reshape(1, { shape_mask.height, shape_mask.width });
#pragma region apply sigmoid to the mask
	cv::exp(-res, res);
	res = 1.0 / (1.0 + res);
#pragma endregion

	cv::resize(res, res, shape_input);

	float scalex = shape_src.width * 1.0 / shape_input.width;
	float scaley = shape_src.height * 1.0 / shape_input.height;
	cv::Mat tmp;
	if (scalex > scaley)
		cv::resize(res, tmp, cv::Size(shape_src.width, static_cast<int>(shape_input.height * scalex)));
	else
		cv::resize(res, tmp, cv::Size(static_cast<int>(shape_input.width * scaley), shape_src.height));

	cv::Mat dst = tmp(cv::Rect(0, 0, shape_src.width, shape_src.height));
	mk = dst(box) > mask_threshold;
	//mk = dst > mask_threshold;
}

void draw_boxes_mask(const std::vector<std::string>& classes, const std::vector<int>& ids, const std::vector<float>& confidences,
	const std::vector<cv::Rect>& boxes, const std::vector<cv::Mat>& masks, const std::string& name, cv::Mat& frame, const std::string& result_dir)
{
	std::cout << "image name: " << name << ", number of detections: " << ids.size() << std::endl;

	std::random_device rd;
	std::mt19937 gen(rd());
	std::uniform_int_distribution<int> dis(100, 255);
	cv::Mat mk = frame.clone();
	cv::Mat mask = cv::Mat::zeros(mk.size(), CV_8UC1);

	std::vector<cv::Scalar> colors;
	for (auto i = 0; i < classes.size(); ++i)
		colors.emplace_back(cv::Scalar(dis(gen), dis(gen), dis(gen)));

	for (auto i = 0; i < ids.size(); ++i) {

#if 0 可以打开
		cv::rectangle(frame, boxes[i], colors[ids[i]], 2);

		std::string class_string = classes[ids[i]] + ' ' + std::to_string(confidences[i]).substr(0, 4);
		cv::Size text_size = cv::getTextSize(class_string, cv::FONT_HERSHEY_DUPLEX, 1, 2, 0);
		cv::Rect text_box(boxes[i].x, boxes[i].y - 40, text_size.width + 10, text_size.height + 20);

		cv::rectangle(frame, text_box, colors[ids[i]], cv::FILLED);
		cv::putText(frame, class_string, cv::Point(boxes[i].x + 5, boxes[i].y - 10), cv::FONT_HERSHEY_DUPLEX, 1, cv::Scalar(0, 0, 0), 2, 0);
#endif

		mk(boxes[i]).setTo(colors[ids[i]], masks[i]);
		mask(boxes[i]).setTo(cv::Scalar(255), masks[i]);
	}

	cv::addWeighted(frame, 0.5, mk, 0.5, 0, frame);

	// 显示mask
	cv::imshow("Inference", mask);
	cv::waitKey(-1);

	cv::imwrite(result_dir + "/" + name, frame);
}


int main()
{
	namespace fs = std::filesystem;

	const char* onnx_file{ "D:/cppProj/YOLOV8InferTest/best.onnx" };
	constexpr int input_size[2]{ 640, 640 };
	constexpr float confidence_threshold{ 0.45 };
	constexpr float iou_threshold{ 0.50 }; // iou threshold
	constexpr float mask_threshold{ 0.50 }; // segment mask threshold

	constexpr bool cuda_enabled{ false };
	const char* classes_file{ "D:/cppProj/YOLOV8InferTest/classes.txt" };
	const std::string testPath("D:/pythonProj/torch_projs/test/catdog/000012.jpg");
	// 分离出文件名
	std::string filename;
	if (std::filesystem::exists(std::filesystem::path(testPath))==false)
	{
		std::cout << testPath << " is not exists." << std::endl;
		return -1;		
	}
	else
	{
		filename = std::filesystem::path(testPath).filename().string();
		std::cout << "filename = " << filename << std::endl;
	}

	const char* result_dir{ "./predictResult" };

	auto net = cv::dnn::readNetFromONNX(onnx_file);
	if (net.empty()) 
	{
		std::cerr << "Error: there are no layers in the network: " << onnx_file << std::endl;
		return -1;
	}
	
	if (cuda_enabled) {
		net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
		net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);
	}
	else {
		net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
		net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
	}

	if (!fs::exists(result_dir)) 
	{
		fs::create_directories(result_dir);
	}

	auto classes = parse_classes_file(classes_file);
	if (classes.size() == 0) 
	{
		std::cerr << "Error: fail to parse classes file: " << classes_file << std::endl;
		return -1;
	}

	std::cout << "classes: ";
	for (const auto& val : classes) {
		std::cout << val << " ";
	}
	std::cout << std::endl;

#pragma region 预测
	cv::Mat frame = cv::imread(testPath, cv::IMREAD_COLOR);
	if (frame.empty()) 
	{
		std::cerr << "Warning: unable to load image: " << testPath << std::endl;
		return -1;
	}

	auto tstart = std::chrono::high_resolution_clock::now();
	cv::Mat bgr = modify_image_size(frame);
	std::cout << "bgr.size=" << bgr.size << std::endl;

	cv::Mat blob;
	cv::dnn::blobFromImage(bgr, blob, 1.0 / 255.0, cv::Size(input_size[1], input_size[0]), cv::Scalar(), true, false);
	net.setInput(blob);

	std::vector<cv::Mat> outputs;
	net.forward(outputs, net.getUnconnectedOutLayersNames());
	if (outputs.size() != 2) 
	{
		std::cerr << "Error: output must have 2 layers: " << outputs.size() << std::endl;
		return -1;
	}
	else
	{
		std::cout << "outputs.size() = " << outputs.size() << std::endl;
	}

	std::cout << "outputs[0].size=" << outputs[0].size << std::endl;

	// output0
	cv::Mat data0 = cv::Mat(outputs[0].size[1], outputs[0].size[2], CV_32FC1, outputs[0].data).t();
	std::cout << "data0.size = " << data0.size << std::endl;
	
	std::cout << outputs[0].size << std::endl;

	// output1
	std::vector<int> sizes;
	std::cout << outputs[1].size << std::endl;

	for (int i = 0; i < 4; ++i)
	{
		sizes.emplace_back(outputs[1].size[i]);
		std::cout << sizes[i] << " ";
	}
	std::cout << std::endl;

	cv::Mat data1 = cv::Mat(sizes, CV_32F, outputs[1].data);
	std::cout << "data1.size = " << data1.size << std::endl;

	auto tend = std::chrono::high_resolution_clock::now();
	std::cout << "elapsed millisenconds: " << std::chrono::duration_cast<std::chrono::milliseconds>(tend - tstart).count() << " ms" << std::endl;
#pragma endregion

#pragma region 后处理
	std::vector<int> class_ids;
	std::vector<float> confidences;
	std::vector<cv::Rect> boxes;
	std::vector<std::vector<float>> masks;

	float scalex = frame.cols * 1.f / input_size[1]; // note: image_preprocess function
	float scaley = frame.rows * 1.f / input_size[0];
	auto scale = (scalex > scaley) ? scalex : scaley;

	const float* data = (float*)data0.data;
	for (auto i = 0; i < data0.rows; ++i)
	{
		cv::Mat scores(1, classes.size(), CV_32FC1, (float*)data + 4);
		//std::cout << "scores.size=" << scores.size << std::endl;
		

		cv::Point class_id;
		double max_class_score;

		cv::minMaxLoc(scores, 0, &max_class_score, 0, &class_id);
		//std::cout << "classid.x=" << class_id.x << std::endl;
		

		if (max_class_score > confidence_threshold) 
		{
			confidences.emplace_back(max_class_score);
			class_ids.emplace_back(class_id.x);
			// 大小为data[6]
			
			// 取出data[6,cols)的内容
			masks.emplace_back(
				std::vector<float>(data + 4 + classes.size(), data + data0.cols)
			); // 32		

			float x = data[0];
			float y = data[1];
			float w = data[2];
			float h = data[3];

			int left = std::max(0, std::min(int((x - 0.5 * w) * scale), frame.cols));
			int top = std::max(0, std::min(int((y - 0.5 * h) * scale), frame.rows));
			int width = std::max(0, std::min(int(w * scale), frame.cols - left));
			int height = std::max(0, std::min(int(h * scale), frame.rows - top));
			boxes.emplace_back(cv::Rect(left, top, width, height));
		}
		data += data0.cols;
	}
	std::cout << "masks.size=" << masks.size() << std::endl;

	std::vector<int> nms_result;// 返回得分最高的box的索引
	cv::dnn::NMSBoxes(boxes, confidences, confidence_threshold, iou_threshold, nms_result);

	// 0表示元素总个数不变
	cv::Mat proto = data1.reshape(0, { sizes[1], sizes[2] * sizes[3] });

	std::vector<int> ids;
	std::vector<float> confs;
	std::vector<cv::Rect> rects;
	std::vector<cv::Mat> mks;
	for (size_t i = 0; i < nms_result.size(); ++i) 
	{
		auto index = nms_result[i];
		std::cout << "index=" << index << std::endl;

		ids.emplace_back(class_ids[index]);
		confs.emplace_back(confidences[index]);
		boxes[index] = boxes[index] & cv::Rect(0, 0, frame.cols, frame.rows);
		
		// 原始大小
		std::cout << "cols=" << frame.cols << ", rows=" << frame.rows << std::endl;
		
		std::cout << "cv::Mat(masks[index]).size=" << cv::Mat(masks[index]).size << std::endl;

		cv::Mat mk;
		get_masks(
			cv::Mat(masks[index]).t(), proto, sizes, frame, boxes[index], mk, input_size, mask_threshold);
		mks.emplace_back(mk);
		rects.emplace_back(boxes[index]);

		/*std::cout << "ids=" << ids[i] << std::endl;
		cv::imshow("w", mk);
		cv::waitKey();*/
	}
	draw_boxes_mask(classes, ids, confs, rects, mks, filename, frame, result_dir);

#pragma endregion


	return 0;
}