c++调用tensorflow object detect api 生成的模型

//opencv-4.2.0 tensorflow 1.14.0 亲测可用

#include <fstream>
#include <sstream>
#include <opencv2/dnn.hpp>
#include <opencv2/imgproc.hpp>
#include <opencv2/highgui.hpp>
#include <iostream>
using namespace cv;
using namespace dnn;
using namespace std;
float confThreshold, nmsThreshold;
std::vector<std::string> classes;
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);
int main(int argc, char** argv)
{
    // 根据选择的检测模型文件进行配置
    confThreshold = 0.5;
    nmsThreshold = 0.4;
    float scale = 1.0;
    Scalar mean = { 0, 0, 0 };
    bool swapRB = true;
    int inpWidth = 300;
    int inpHeight = 300;
    // String weights = "/home/ssd_mobilenet_v1_coco_11_06_2017/frozen_inference_graph.pb";
        //String prototxt = "/home/opencv_extra-master/testdata/dnn/ssd_mobilenet_v1_coco.pbtxt";
    //String modelPath = "/home/ssd_mobilenet_v1_coco_11_06_2017/frozen_inference_graph.pb";
    String modelPath = "/home/cctv_output_dir/frozen_inference_graph.pb";

//　　这个cctv.pbtxt 是通过opencv源码/sample/dnn下的一个脚本生成的
    String configPath = "/home/cctv_output_dir/cctv.pbtxt";
    String framework = "";
    int backendId = cv::dnn::DNN_BACKEND_OPENCV;
    int targetId = cv::dnn::DNN_TARGET_CPU;
    String classesFile = R"(/root/opencv_src/opencv-4.2.0/samples/data/dnn/object_detection_classes_coco.txt)";
    // Open file with classes names.
    if (!classesFile.empty()) {
        const std::string& file = classesFile;
        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, framework);

    net.setPreferableBackend(backendId);

    net.setPreferableTarget(targetId);

    std::vector<String> outNames = net.getUnconnectedOutLayersNames();

    // Create a window

    static const std::string kWinName = "Deep learning object detection in OpenCV";

    // Process frames.

    Mat frame, blob;

    frame = imread("/root/cctv_test_image/cctv10.jpg");

    // Create a 4D blob from a frame.

    Size inpSize(inpWidth > 0 ? inpWidth : frame.cols,
        inpHeight > 0 ? inpHeight : frame.rows);

    blobFromImage(frame, blob, scale, inpSize, mean, swapRB, false);

    // Run a model.

    net.setInput(blob);

    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");
    }

    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);
    waitKey(0);
    return 0;
}

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;
    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 > 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;
                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() == 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 > confThreshold) {
                int left = (int)(data[i + 3] * frame.cols);
                int top = (int)(data[i + 4] * frame.rows);
                int right = (int)(data[i + 5] * frame.cols);
                int bottom = (int)(data[i + 6] * frame.rows);
                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 == "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());
}

效果图：