====== OpenCV DNN with CUDA ======
このページでは，OpenCVのdnnモジュールをcudaでinferenceさせるための環境構築に関してまとめます．もともとの動機は
  * OpenCVのdnn inferenceをもっと早くしたい
なわけです．もちろんintelのinference engine ( https://github.com/opencv/opencv/wiki/Intel%27s-Deep-Learning-Inference-Engine-backend )を利用するのもありですが，導入したところで1.5倍程度の速度向上しか見込めません．若干モチベーションが下がる．それで，dnn moduleのcudaサポートがついに4.2から実現した（対応ネットワーク構成に制限があります）という情報を掴み，早速cudaでyoloやssd等を走らせてみよう．と思った次第です．実際にcudaで実行してあげると，大体CPU inferenceの15倍位（Geforce GTX 1080Tiのとき）になります．さすがcuda．

参考にした記事は以下となります．日本語でこのあたりをubuntu環境でやってる人がいなかったのでここに記しておくことにしました．
  * https://www.pyimagesearch.com/2020/02/03/how-to-use-opencvs-dnn-module-with-nvidia-gpus-cuda-and-cudnn/
  * https://gist.github.com/ollewelin/4d9164fa30a9403e87c2ad193655f42a
  * https://gist.github.com/YashasSamaga/985071dc57885348bec072b4dc23824f
    * dnn with cudaを実装した人であるYashasSamagaさんがissueで色々質問に答えています．

基本的には上記のリンクに書いてある通りにすればコンパイルできるんでないかなと思います．

===== 環境構築 =====
  * Ubuntu 16.04
  * CUDA Toolkit 10.2
  * cuDNN 7.6.4

cmakeでconfigureする際に，導入しているバージョンが複数ある場合は，適切なバージョンに変更するなどのマニュアル作業が生じます．できればcmake-guiを利用してパスやバージョンが正しいかを細かく確認することをおすすめします．特にmakeに難しいことはないですが，エラーが出る場合はcuda周りを一通りチェックしてください．また，opencv_contribが必要なのでそれも忘れずに．必要に応じてnvidia-dockerすると良いかなと思います．

==== object_detection.cppの修正 ====
デフォルトのobject_detection.cppが若干古いのと扱いづらいところがあるので，以下のものと入れ替えてください．
参照パスは　opencv/samples/dnn/object_detection.cpp です．

<file .cpp object_detection.cpp>
#include <fstream>
#include <sstream>

#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>

#ifdef CV_CXX11
#include <mutex>
#include <thread>
#include <queue>
#endif

#include "common.hpp"

std::string keys =
    "{ help  h     | | Print help message. }"
    "{ @alias      | | An alias name of model to extract preprocessing parameters from models.yml file. }"
    "{ zoo         | models.yml | An optional path to file with preprocessing parameters }"
    "{ device      |  0 | camera device number. }"
    "{ input i     | | Path to input image or video file. Skip this argument to capture frames from a camera. }"
    "{ framework f | | Optional name of an origin framework of the model. Detect it automatically if it does not set. }"
    "{ classes     | | Optional path to a text file with names of classes to label detected objects. }"
    "{ thr         | .5 | Confidence threshold. }"
    "{ nms         | .4 | Non-maximum suppression threshold. }"
    "{ backend     |  0 | Choose one of computation backends: "
                         "0: automatically (by default), "
                         "1: Halide language (http://halide-lang.org/), "
                         "2: Intel's Deep Learning Inference Engine (https://software.intel.com/openvino-toolkit), "
                         "3: OpenCV implementation,"
                         "4: VKCOM,"
                         "5:CUDA"
                         "}"
    "{ target      | 0 | Choose one of target computation devices: "
                         "0: CPU target (by default), "
                         "1: OpenCL, "
                         "2: OpenCL fp16 (half-float precision), "
                         "3: VPU,"
                         "4:VULKANm,"
                         "5:FPGA,"
                         "6:CUDA,"
                         "7:CUDA_FP16"
                         "}"
    "{ async       | 0 | Number of asynchronous forwards at the same time. "
                        "Choose 0 for synchronous mode }";

using namespace cv;
using namespace dnn;

float confThreshold, nmsThreshold;
std::vector<std::string> classes;

inline void preprocess(const Mat& frame, Net& net, Size inpSize, float scale,
                       const Scalar& mean, bool swapRB);

void postprocess(Mat& frame, const std::vector<Mat>& out, Net& net);

void drawPred(int classId, float conf, int left, int top, int right, int bottom, Mat& frame);

void callback(int pos, void* userdata);

#ifdef CV_CXX11
template <typename T>
class QueueFPS : public std::queue<T>
{
public:
    QueueFPS() : counter(0) {}

    void push(const T& entry)
    {
        std::lock_guard<std::mutex> lock(mutex);

        std::queue<T>::push(entry);
        counter += 1;
        if (counter == 1)
        {
            // Start counting from a second frame (warmup).
            tm.reset();
            tm.start();
        }
    }

    T get()
    {
        std::lock_guard<std::mutex> lock(mutex);
        T entry = this->front();
        this->pop();
        return entry;
    }

    float getFPS()
    {
        tm.stop();
        double fps = counter / tm.getTimeSec();
        tm.start();
        return static_cast<float>(fps);
    }

    void clear()
    {
        std::lock_guard<std::mutex> lock(mutex);
        while (!this->empty())
            this->pop();
    }

    unsigned int counter;

private:
    TickMeter tm;
    std::mutex mutex;
};
#endif  // CV_CXX11

int main(int argc, char** argv)
{
    CommandLineParser parser(argc, argv, keys);

    const std::string modelName = parser.get<String>("@alias");
    const std::string zooFile = parser.get<String>("zoo");

    keys += genPreprocArguments(modelName, zooFile);

    parser = CommandLineParser(argc, argv, keys);
    parser.about("Use this script to run object detection deep learning networks using OpenCV.");
    if (argc == 1 || parser.has("help"))
    {
        parser.printMessage();
        return 0;
    }

    confThreshold = parser.get<float>("thr");
    nmsThreshold = parser.get<float>("nms");
    float scale = parser.get<float>("scale");
    Scalar mean = parser.get<Scalar>("mean");
    bool swapRB = parser.get<bool>("rgb");
    int inpWidth = parser.get<int>("width");
    int inpHeight = parser.get<int>("height");
    size_t asyncNumReq = parser.get<int>("async");
    CV_Assert(parser.has("model"));
    std::string modelPath = findFile(parser.get<String>("model"));
    std::string configPath = findFile(parser.get<String>("config"));

    // Open file with classes names.
    if (parser.has("classes"))
    {
        std::string file = parser.get<String>("classes");
        std::ifstream ifs(file.c_str());
        if (!ifs.is_open())
            CV_Error(Error::StsError, "File " + file + " not found");
        std::string line;
        while (std::getline(ifs, line))
        {
            classes.push_back(line);
        }
    }

    // Load a model.
    Net net = readNet(modelPath, configPath, parser.get<String>("framework"));
    net.setPreferableBackend(parser.get<int>("backend"));
    net.setPreferableTarget(parser.get<int>("target"));
    std::vector<String> outNames = net.getUnconnectedOutLayersNames();

    // Create a window
    static const std::string kWinName = "Deep learning object detection in OpenCV";
    namedWindow(kWinName, WINDOW_NORMAL);
    int initialConf = (int)(confThreshold * 100);
    createTrackbar("Confidence threshold, %", kWinName, &initialConf, 99, callback);

    // Open a video file or an image file or a camera stream.
    VideoCapture cap;
    if (parser.has("input"))
        cap.open(parser.get<String>("input"));
    else
        cap.open(parser.get<int>("device"));

#ifdef CV_CXX11
    bool process = true;

    // Frames capturing thread
    QueueFPS<Mat> framesQueue;
    std::thread framesThread([&](){
        Mat frame;
        while (process)
        {
            cap >> frame;
            if (!frame.empty())
                framesQueue.push(frame.clone());
            else
                break;
        }
    });

    // Frames processing thread
    QueueFPS<Mat> processedFramesQueue;
    QueueFPS<std::vector<Mat> > predictionsQueue;
    std::thread processingThread([&](){
        std::queue<AsyncArray> futureOutputs;
        Mat blob;
        while (process)
        {
            // Get a next frame
            Mat frame;
            {
                if (!framesQueue.empty())
                {
                    frame = framesQueue.get();
                    if (asyncNumReq)
                    {
                        if (futureOutputs.size() == asyncNumReq)
                            frame = Mat();
                    }
                    else
                        framesQueue.clear();  // Skip the rest of frames
                }
            }

            // Process the frame
            if (!frame.empty())
            {
                preprocess(frame, net, Size(inpWidth, inpHeight), scale, mean, swapRB);
                processedFramesQueue.push(frame);

                if (asyncNumReq)
                {
                    futureOutputs.push(net.forwardAsync());
                }
                else
                {
                    std::vector<Mat> outs;
                    net.forward(outs, outNames);
                    predictionsQueue.push(outs);
                }
            }

            while (!futureOutputs.empty() &&
                   futureOutputs.front().wait_for(std::chrono::seconds(0)))
            {
                AsyncArray async_out = futureOutputs.front();
                futureOutputs.pop();
                Mat out;
                async_out.get(out);
                predictionsQueue.push({out});
            }
        }
    });

    // Postprocessing and rendering loop
    while (waitKey(1) < 0)
    {
        if (predictionsQueue.empty())
            continue;

        std::vector<Mat> outs = predictionsQueue.get();
        Mat frame = processedFramesQueue.get();

        postprocess(frame, outs, net);

        //if (predictionsQueue.counter > 1)
        //{
            std::string label = format("Camera: %.2f FPS", framesQueue.getFPS());
            putText(frame, label, Point(0, 15), FONT_HERSHEY_DUPLEX, 0.5, Scalar(0, 0, 0));

            label = format("Network: %.2f FPS", predictionsQueue.getFPS());
            putText(frame, label, Point(0, 30), FONT_HERSHEY_DUPLEX, 0.5, Scalar(0, 0, 0));

            label = format("Skipped frames: %d", framesQueue.counter - predictionsQueue.counter);
            putText(frame, label, Point(0, 45), FONT_HERSHEY_DUPLEX, 0.5, Scalar(0, 0, 0));

            std::vector<double> layersTimes;
            double freq = getTickFrequency() / 1000;
            double t = net.getPerfProfile(layersTimes) / freq;
            label = format("Inference time: %.2f ms", t);
            putText(frame, label, Point(0, 60), FONT_HERSHEY_DUPLEX, 0.5, Scalar(0, 0, 0));
        //}
        imshow(kWinName, frame);
    }

    process = false;
    framesThread.join();
    processingThread.join();

#else  // CV_CXX11
    if (asyncNumReq)
        CV_Error(Error::StsNotImplemented, "Asynchronous forward is supported only with Inference Engine backend.");

    // Process frames.
    Mat frame, blob;
    while (waitKey(1) < 0)
    {
        cap >> frame;
        if (frame.empty())
        {
            waitKey();
            break;
        }

        preprocess(frame, net, Size(inpWidth, inpHeight), scale, mean, swapRB);

        std::vector<Mat> outs;
        net.forward(outs, outNames);

        postprocess(frame, outs, net);

        // Put efficiency information.
        std::vector<double> layersTimes;
        double freq = getTickFrequency() / 1000;
        double t = net.getPerfProfile(layersTimes) / freq;
        std::string label = format("Inference time: %.2f ms", t);
        putText(frame, label, Point(0, 15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0));

        imshow(kWinName, frame);
    }
#endif  // CV_CXX11
    return 0;
}

inline void preprocess(const Mat& frame, Net& net, Size inpSize, float scale,
                       const Scalar& mean, bool swapRB)
{
    static Mat blob;
    // Create a 4D blob from a frame.
    if (inpSize.width <= 0) inpSize.width = frame.cols;
    if (inpSize.height <= 0) inpSize.height = frame.rows;
    blobFromImage(frame, blob, 1.0, inpSize, Scalar(), swapRB, false, CV_8U);

    // Run a model.
    net.setInput(blob, "", scale, mean);
    if (net.getLayer(0)->outputNameToIndex("im_info") != -1)  // Faster-RCNN or R-FCN
    {
        resize(frame, frame, inpSize);
        Mat imInfo = (Mat_<float>(1, 3) << inpSize.height, inpSize.width, 1.6f);
        net.setInput(imInfo, "im_info");
    }
}

void postprocess(Mat& frame, const std::vector<Mat>& outs, Net& net)
{
    static std::vector<int> outLayers = net.getUnconnectedOutLayers();
    static std::string outLayerType = net.getLayer(outLayers[0])->type;

    std::vector<int> classIds;
    std::vector<float> confidences;
    std::vector<Rect> boxes;
    boxes.clear();
     if (net.getLayer(0)->outputNameToIndex("im_info") != -1)  // Faster-RCNN or R-FCN
    {
        /*
        // Network produces output blob with a shape 1x1xNx7 where N is a number of
        // detections and an every detection is a vector of values
        // [batchId, classId, confidence, left, top, right, bottom]
        CV_Assert(outs.size() == 1);
        float* data = (float*)outs[0].data;
        for (size_t i = 0; i < outs[0].total(); i += 7)
        {
            float confidence = data[i + 2];
            if (confidence > confidenceThreshold)
            {
                int left = (int)data[i + 3];
                int top = (int)data[i + 4];
                int right = (int)data[i + 5];
                int bottom = (int)data[i + 6];
                int width = right - left + 1;
                int height = bottom - top + 1;
                classIds.push_back((int)(data[i + 1]) - 1);  // Skip 0th background class id.
//                boxes.push_back(Rect(left, top, width, height));
                confidences.push_back(confidence);
            }
        }
        */
    }
    else if (outLayerType == "DetectionOutput")
    {
        // Network produces output blob with a shape 1x1xNx7 where N is a number of
        // detections and an every detection is a vector of values
        // [batchId, classId, confidence, left, top, right, bottom]
        CV_Assert(outs.size() > 0);
        for (size_t k = 0; k < outs.size(); k++)
        {
            float* data = (float*)outs[k].data;
            for (size_t i = 0; i < outs[k].total(); i += 7)
            {
                float confidence = data[i + 2];
                if (confidence > confThreshold)
                {
                    int left   = (int)data[i + 3];
                    int top    = (int)data[i + 4];
                    int right  = (int)data[i + 5];
                    int bottom = (int)data[i + 6];
                    int width  = right - left + 1;
                    int height = bottom - top + 1;
                    if (width <= 2 || height <= 2)
                    {
                        left   = (int)(data[i + 3] * frame.cols);
                        top    = (int)(data[i + 4] * frame.rows);
                        right  = (int)(data[i + 5] * frame.cols);
                        bottom = (int)(data[i + 6] * frame.rows);
                        width  = right - left + 1;
                        height = bottom - top + 1;
                    }
                    classIds.push_back((int)(data[i + 1]) - 1);  // Skip 0th background class id.
                    boxes.push_back(Rect(left, top, width, height));
                    confidences.push_back(confidence);
                }
            }
        }
    }
    else if (outLayerType == "Region")
    {
        for (size_t i = 0; i < outs.size(); ++i)
        {
            // Network produces output blob with a shape NxC where N is a number of
            // detected objects and C is a number of classes + 4 where the first 4
            // numbers are [center_x, center_y, width, height]
            float* data = (float*)outs[i].data;
            for (int j = 0; j < outs[i].rows; ++j, data += outs[i].cols)
            {
                Mat scores = outs[i].row(j).colRange(5, outs[i].cols);
                Point classIdPoint;
                double confidence;
                minMaxLoc(scores, 0, &confidence, 0, &classIdPoint);
                if (confidence > confThreshold)
                {
                    int centerX = (int)(data[0] * frame.cols);
                    int centerY = (int)(data[1] * frame.rows);
                    int width = (int)(data[2] * frame.cols);
                    int height = (int)(data[3] * frame.rows);
                    int left = centerX - width / 2;
                    int top = centerY - height / 2;

                    classIds.push_back(classIdPoint.x);
                    confidences.push_back((float)confidence);
                    boxes.push_back(Rect(left, top, width, height));
                }
            }
        }
    }
    else
        CV_Error(Error::StsNotImplemented, "Unknown output layer type: " + outLayerType);


    std::vector<int> indices;
    NMSBoxes(boxes, confidences, confThreshold, nmsThreshold, indices);
    for (size_t i = 0; i < indices.size(); ++i)
    {
        int idx = indices[i];
        Rect box = boxes[idx];
        drawPred(classIds[idx], confidences[idx], box.x, box.y,
                 box.x + box.width, box.y + box.height, frame);
    }
   
   
}

void drawPred(int classId, float conf, int left, int top, int right, int bottom, Mat& frame)
{
    rectangle(frame, Point(left, top), Point(right, bottom), Scalar(0, 255, 0));

    std::string label = format("%.2f", conf);
    if (!classes.empty())
    {
        CV_Assert(classId < (int)classes.size());
        label = classes[classId] + ": " + label;
    }

    int baseLine;
    Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);

    top = max(top, labelSize.height);
    rectangle(frame, Point(left, top - labelSize.height),
              Point(left + labelSize.width, top + baseLine), Scalar::all(255), FILLED);
    putText(frame, label, Point(left, top), FONT_HERSHEY_SIMPLEX, 0.5, Scalar());
}

void callback(int pos, void*)
{
    confThreshold = pos * 0.01f;
}

</file>

===== 実行 =====
backend, targetに関しては下記を参照する．cudaを利用する場合，opencv-4.2.0の場合は backend=5, target=6 を引数で渡すことを忘れないように．
  * https://docs.opencv.org/4.2.0/d6/d0f/group__dnn.html#ga186f7d9bfacac8b0ff2e26e2eab02625