【pytorch】将yolov5模型通过openVINO2022部署至生产环境（二）:C++版本

（一）openvino安装说明及yolov5的onnx生成详见：
【pytorch】将yolov5模型通过openVINO2022部署至生产环境（一）:python版本+异步模式
（二）c++版本代码：
使用2种图片输入读取方式;
读取.onnx模型或XML和bin文件；
yolov5_6vino.h：

#ifndef YOLOV5VINO_H
#define YOLOV5VINO_H
#include <fstream>
#include <opencv2/opencv.hpp>
#include <inference_engine.hpp>
#define NOT_NCS2
using namespace cv;
using namespace dnn;
using namespace std;
using namespace InferenceEngine;

class YOLOVINO
{
public:
    struct Detection
    {
        int class_id;
        float confidence;
        Rect box;
    };
public:
    YOLOVINO();
    ~YOLOVINO();
    void init();
    void loadNet(bool is_cuda);
    Mat formatYolov5(const Mat& source);
    void detect(Mat& image, vector<Detection>& outputs);
    void drawRect(Mat& image, vector<Detection>& outputs);
    void loadClassList();
private:
    //每个类的类别得分
    float m_scoreThreshold = 0.4;
    //大于该值被认为是同一个框：
    float m_nmsThreshold = 0.7;
    //每个框含有物体的概率
    float m_confThreshold = 0.6;

    //"CPU","GPU","MYRIAD"
#ifdef NCS2
    const std::string m_deviceName = "MYRIAD";
    const std::string m_modelFilename = "configFiles/yolov5sNCS2.xml";
#else
    //plugins.xml中的定义值：
    const std::string m_deviceName = "CPU";// "HETERO";
    const std::string m_modelFilename = "yolov5s.onnx";
#endif // NCS2
    const std::string m_classfile = "classes.txt";
    size_t m_numChannels = 0;
    size_t m_inputH = 0;
    size_t m_inputW = 0;
    size_t m_imageSize = 0;
    std::string m_inputName = "";
    std::string m_outputName = "";
    vector<std::string> m_classNames;
    const vector<Scalar> colors = { Scalar(255, 255, 0), Scalar(0, 255, 0), Scalar(0, 255, 255), Scalar(255, 0, 0) };
    InferRequest::Ptr m_inferRequest;
    Blob::Ptr m_inputData;
};
#endif

yolov5_6vino.cpp：

#include "yolov5_6vino.h"
#include "samples/ocv_common.hpp"
//将OpenCV Mat对象中的图像数据传给为InferenceEngine Blob对象
void frameToBlob(const cv::Mat& frame, InferRequest::Ptr& inferRequest, const std::string& inputName) {
    /* 从OpenCV Mat对象中拷贝图像数据到InferenceEngine 输入Blob对象 */
    Blob::Ptr frameBlob = inferRequest->GetBlob(inputName);
    //这个函数方便将8U 0-255转为Blob
    matU8ToBlob<uint8_t>(frame, frameBlob); //uint8_t
    
}
void frameToBlob2(const cv::Mat& frame, InferRequest::Ptr& inferRequest, const std::string& inputName) {
    /* 从OpenCV Mat对象中拷贝图像数据到InferenceEngine 输入Blob对象 */
    Blob::Ptr frameBlob = inferRequest->GetBlob(inputName);
    //为适应yolo特性，将输入值除以255的新函数
    matU8ToBlob2<float>(frame, frameBlob); //uint8_t
}
YOLOVINO::YOLOVINO()
{
    init();
}

YOLOVINO::~YOLOVINO()
{
}

void YOLOVINO::init()
{
    InferenceEngine::Core ie;
    //两种加载形式，直接加载onnx，或者加载
    // 下面这种是直接加载onnx:
    //InferenceEngine::CNNNetwork network = ie.ReadNetwork(m_modelFilename);
    InferenceEngine::CNNNetwork network = ie.ReadNetwork("exported_onnx_model.XML", "exported_onnx_model.bin");
    //输入名及输入指针的数组
    InferenceEngine::InputsDataMap inputs = network.getInputsInfo();
    InferenceEngine::OutputsDataMap outputs = network.getOutputsInfo();
    for (auto item : inputs)
    {
        m_inputName = item.first;
        auto input_data = item.second;
        input_data->setPrecision(Precision::FP32);
        //默认布局与维度有关: C - for 1 - dimensional, NC - for 2 - dimensional, CHW - for 3 - dimensional, NCHW - for 4 - dimensional NCDHW - for 5 - dimensional
        //这里其实无需设置
        input_data->setLayout(Layout::NCHW);
        //内建输入预处理，设置以应对RGB格式的输入
        input_data->getPreProcess().setColorFormat(ColorFormat::RGB);
        //std::cout << "input name = " << m_inputName << std::endl;
    }

    //输出所有输出名：
    for (auto item : outputs)
    {
        auto output_data = item.second;
        output_data->setPrecision(Precision::FP32);
        m_outputName = item.first;
        std::cout << "output name = " << m_outputName << std::endl;
    }
    //加载网络为可执行网络
    auto executable_network = ie.LoadNetwork(network, m_deviceName);
    m_inferRequest = executable_network.CreateInferRequestPtr();

    m_inputData = m_inferRequest->GetBlob(m_inputName);
    m_numChannels = m_inputData->getTensorDesc().getDims()[1];
    m_inputH = m_inputData->getTensorDesc().getDims()[2];
    m_inputW = m_inputData->getTensorDesc().getDims()[3];
    //m_imageSize = m_inputH * m_inputW;

    loadClassList();
}
//该函数将vetor中的数据按行读取
void YOLOVINO::loadClassList()
{
    std::ifstream ifs(m_classfile);
    std::string line;
    while (getline(ifs, line))
        m_classNames.push_back(line);
}
//该函数取图像最大边长作新正方形，多出部分补0
Mat YOLOVINO::formatYolov5(const Mat& source)
{
    int col = source.cols;
    int row = source.rows;
    int max = MAX(col, row);
    Mat result = Mat::zeros(max, max, CV_8UC3);
    source.copyTo(result(Rect(0, 0, col, row)));
    return result;
}


void YOLOVINO::detect(Mat& image, vector<Detection>& outputs)
{
    cv::Mat input_image = formatYolov5(image);
    cv::Mat blob_image;
    //cv::resize(input_image, blob_image, cv::Size(m_inputW, m_inputH));
    cvtColor(input_image, blob_image, COLOR_BGR2RGB);
    //输入图像转换的两种方法：
    //法1：借助安装用例自带的辅助文件samples/ocv_common.hpp
    //frameToBlob(blob_image,  m_inferRequest, m_inputName);
    frameToBlob2(blob_image, m_inferRequest, m_inputName);
    //法2：要先取输入的格式参数：
//    float* data = static_cast<float*>(m_inputData->buffer());
//    for (size_t row = 0; row < m_inputH; row++) {
//        for (size_t col = 0; col < m_inputW; col++) {
//            for (size_t ch = 0; ch < m_numChannels; ch++) {
//#ifdef NCS2
//                data[m_imageSize * ch + row * m_inputW + col] = float(blob_image.at<cv::Vec3b>(row, col)[ch]);
//#else
//                data[m_imageSize * ch + row * m_inputW + col] = float(blob_image.at<cv::Vec3b>(row, col)[ch] / 255.0);
//#endif // NCS2
//            }
//        }
//    }
    auto start = std::chrono::high_resolution_clock::now();

    m_inferRequest->Infer();

    auto output = m_inferRequest->GetBlob(m_outputName);
    const float* detection_out =output->buffer().as<PrecisionTrait<Precision::FP32>::value_type*>();

    //维度信息
    const SizeVector outputDims = output->getTensorDesc().getDims();//1,6300[25200],9
    auto end = std::chrono::high_resolution_clock::now();
    auto time = std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();
    cout << "time = " << time << endl;
    float x_factor = float(input_image.cols / m_inputW);
    float y_factor = float(input_image.rows / m_inputH);
    float* dataout = (float*)detection_out;
    const int dimensions = outputDims[2];
    const int rows = outputDims[1];
    vector<int> class_ids;
    vector<float> confidences;
    vector<Rect> boxes;
    //解析顺序，维度为1的可直接略掉。
    //25200个9依次排列：
    for (int i = 0; i < rows; ++i)
    {
        float confidence = dataout[4];
        if (confidence >= m_confThreshold)
        {
            float* classes_scores = dataout + 5;
            Mat scores(1, m_classNames.size(), CV_32FC1, classes_scores);
            Point class_id;
            double max_class_score;
            minMaxLoc(scores, 0, &max_class_score, 0, &class_id);
            if (max_class_score > m_scoreThreshold)
            {
                confidences.push_back(confidence);
                class_ids.push_back(class_id.x);

                float x = dataout[0];
                float y = dataout[1];
                float w = dataout[2];
                float h = dataout[3];
                int left = int((x - 0.5 * w) * x_factor);
                int top = int((y - 0.5 * h) * y_factor);
                int width = int(w * x_factor);
                int height = int(h * y_factor);
                boxes.push_back(Rect(left, top, width, height));
            }
        }
        dataout += dimensions;
    }
    vector<int> nms_result;
    NMSBoxes(boxes, confidences, m_scoreThreshold, m_nmsThreshold, nms_result);
    for (int i = 0; i < nms_result.size(); i++)
    {
        int idx = nms_result[i];
        Detection result;
        result.class_id = class_ids[idx];
        result.confidence = confidences[idx];
        result.box = boxes[idx];
        outputs.push_back(result);
    }
}
//绘制box框
void YOLOVINO::drawRect(Mat& image, vector<Detection>& outputs)
{
    int detections = outputs.size();
    for (int i = 0; i < detections; ++i)
    {
        auto detection = outputs[i];
        auto box = detection.box;
        auto classId = detection.class_id;
        const auto color = colors[classId % colors.size()];
        rectangle(image, box, color, 3);

        rectangle(image, Point(box.x, box.y - 40), Point(box.x + box.width, box.y), color, FILLED);
        putText(image, m_classNames[classId].c_str(), Point(box.x, box.y - 5), FONT_HERSHEY_SIMPLEX, 1.5, Scalar(0, 0, 0), 2);
    }

}

main.cpp：

#include "inference_engine.hpp"
#include "opencv2/opencv.hpp"
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include "yolov5_6vino.h"
using namespace std;
using namespace cv;
using namespace dnn;
using namespace InferenceEngine;

int main(int argc, char** argv)
{
	YOLOVINO yolov5vino;
	int frameCount = 0;
	int fps = 0;
	Mat frame;

	auto start = std::chrono::high_resolution_clock::now();
	VideoCapture capture("sample.mp4");
	while (1)
	{
		capture.read(frame);
		if (frame.empty())
		{
			break;
		}
		++frameCount;
		//灰度图像也按RGB处理
		if (frame.channels() < 3)
			cvtColor(frame, frame, COLOR_GRAY2RGB);

		std::vector<YOLOVINO::Detection> outputs;
		yolov5vino.detect(frame, outputs);
		yolov5vino.drawRect(frame, outputs);
		if (frameCount >= 15)
		{

			auto end = std::chrono::high_resolution_clock::now();
			//计算fps：
			fps = frameCount * 1000.0 / std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();

			frameCount = 0;
			start = std::chrono::high_resolution_clock::now();
		}

		if (fps > 0)
		{

			std::ostringstream fps_label;
			fps_label << std::fixed << std::setprecision(2);
			fps_label << "FPS: " << fps;
			std::string fps_label_str = fps_label.str();
			cv::putText(frame, fps_label_str.c_str(), cv::Point(20, 50), cv::FONT_HERSHEY_SIMPLEX, 2, cv::Scalar(0, 0, 255), 2);
		}
		cv::namedWindow("output", cv::WINDOW_NORMAL);
		cv::imshow("output", frame);

		if (cv::waitKey(1) != -1)
		{
			capture.release();
			std::cout << "finished by user\n";
			break;
		}
	}
	return 1;
}

yolo的输入需要在原图片基础上除以255，更新后的辅助函数文件为：

// Copyright (C) 2018-2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

/**
 * @brief a header file with common samples functionality using OpenCV
 * @file ocv_common.hpp
 */

#pragma once

#include <opencv2/opencv.hpp>

#include "openvino/openvino.hpp"
#include "samples/common.hpp"

/**
 * @brief Sets image data stored in cv::Mat object to a given Blob object.
 * @param orig_image - given cv::Mat object with an image data.
 * @param blob - Blob object which to be filled by an image data.
 * @param batchIndex - batch index of an image inside of the blob.
 */
template <typename T>
void matU8ToBlob(const cv::Mat& orig_image, InferenceEngine::Blob::Ptr& blob, int batchIndex = 0) {
    InferenceEngine::SizeVector blobSize = blob->getTensorDesc().getDims();
    const size_t width = blobSize[3];
    const size_t height = blobSize[2];
    const size_t channels = blobSize[1];
    InferenceEngine::MemoryBlob::Ptr mblob = InferenceEngine::as<InferenceEngine::MemoryBlob>(blob);
    if (!mblob) {
        IE_THROW() << "We expect blob to be inherited from MemoryBlob in matU8ToBlob, "
                   << "but by fact we were not able to cast inputBlob to MemoryBlob";
    }
    // locked memory holder should be alive all time while access to its buffer happens
    auto mblobHolder = mblob->wmap();

    T* blob_data = mblobHolder.as<T*>();

    cv::Mat resized_image(orig_image);
    if (static_cast<int>(width) != orig_image.size().width || static_cast<int>(height) != orig_image.size().height) {
        cv::resize(orig_image, resized_image, cv::Size(width, height));
    }

    int batchOffset = batchIndex * width * height * channels;

    for (size_t c = 0; c < channels; c++) {
        for (size_t h = 0; h < height; h++) {
            for (size_t w = 0; w < width; w++) {
                blob_data[batchOffset + c * width * height + h * width + w] = resized_image.at<cv::Vec3b>(h, w)[c];
            }
        }
    }
}
template <typename T>
void matU8ToBlob2(const cv::Mat& orig_image, InferenceEngine::Blob::Ptr& blob, int batchIndex = 0) {
    InferenceEngine::SizeVector blobSize = blob->getTensorDesc().getDims();
    const size_t width = blobSize[3];
    const size_t height = blobSize[2];
    const size_t channels = blobSize[1];
    InferenceEngine::MemoryBlob::Ptr mblob = InferenceEngine::as<InferenceEngine::MemoryBlob>(blob);
    if (!mblob) {
        IE_THROW() << "We expect blob to be inherited from MemoryBlob in matU8ToBlob, "
            << "but by fact we were not able to cast inputBlob to MemoryBlob";
    }
    // locked memory holder should be alive all time while access to its buffer happens
    auto mblobHolder = mblob->wmap();

    T* blob_data = mblobHolder.as<T*>();

    cv::Mat resized_image(orig_image);
    if (static_cast<int>(width) != orig_image.size().width || static_cast<int>(height) != orig_image.size().height) {
        cv::resize(orig_image, resized_image, cv::Size(width, height));
    }

    int batchOffset = batchIndex * width * height * channels;

    for (size_t c = 0; c < channels; c++) {
        for (size_t h = 0; h < height; h++) {
            for (size_t w = 0; w < width; w++) {
                blob_data[batchOffset + c * width * height + h * width + w] = resized_image.at<cv::Vec3b>(h, w)[c]/255.0;
            }
        }
    }
}
/**
 * @brief Wraps data stored inside of a passed cv::Mat object by new Blob pointer.
 * @note: No memory allocation is happened. The blob just points to already existing
 *        cv::Mat data.
 * @param mat - given cv::Mat object with an image data.
 * @return resulting Blob pointer.
 */
static UNUSED InferenceEngine::Blob::Ptr wrapMat2Blob(const cv::Mat& mat) {
    size_t channels = mat.channels();
    size_t height = mat.size().height;
    size_t width = mat.size().width;

    size_t strideH = mat.step.buf[0];
    size_t strideW = mat.step.buf[1];

    bool is_dense = strideW == channels && strideH == channels * width;

    if (!is_dense)
        IE_THROW() << "Doesn't support conversion from not dense cv::Mat";

    InferenceEngine::TensorDesc tDesc(InferenceEngine::Precision::U8,
                                      {1, channels, height, width},
                                      InferenceEngine::Layout::NHWC);

    return InferenceEngine::make_shared_blob<uint8_t>(tDesc, mat.data);
}

static UNUSED ov::Tensor wrapMat2Tensor(const cv::Mat& mat) {
    const size_t channels = mat.channels();
    const size_t height = mat.size().height;
    const size_t width = mat.size().width;

    const size_t strideH = mat.step.buf[0];
    const size_t strideW = mat.step.buf[1];

    const bool is_dense = strideW == channels && strideH == channels * width;
    OPENVINO_ASSERT(is_dense, "Doesn't support conversion from not dense cv::Mat");

    return ov::Tensor(ov::element::u8, ov::Shape{1, height, width, channels}, mat.data);
}

最终测试效果：