vpDetectorDNNOpenCV.cpp

/*
 * ViSP, open source Visual Servoing Platform software.
 * Copyright (C) 2005 - 2024 by Inria. All rights reserved.
 *
 * This software is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 * See the file LICENSE.txt at the root directory of this source
 * distribution for additional information about the GNU GPL.
 *
 * For using ViSP with software that can not be combined with the GNU
 * GPL, please contact Inria about acquiring a ViSP Professional
 * Edition License.
 *
 * See https://visp.inria.fr for more information.
 *
 * This software was developed at:
 * Inria Rennes - Bretagne Atlantique
 * Campus Universitaire de Beaulieu
 * 35042 Rennes Cedex
 * France
 *
 * If you have questions regarding the use of this file, please contact
 * Inria at visp@inria.fr
 *
 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
 *
 * Description:
 * DNN object detection using OpenCV DNN module.
 */
 
#include <visp3/core/vpConfig.h>
 
// Check if std:c++17 or higher
#if (VISP_HAVE_OPENCV_VERSION >= 0x030403) && defined(HAVE_OPENCV_DNN) && \
    ((__cplusplus >= 201703L) || (defined(_MSVC_LANG) && (_MSVC_LANG >= 201703L)))
 
#include <visp3/core/vpImageConvert.h>
#include <visp3/detection/vpDetectorDNNOpenCV.h>
#include <visp3/core/vpIoTools.h>
 
#include<algorithm>
 
BEGIN_VISP_NAMESPACE
  std::string vpDetectorDNNOpenCV::getAvailableDnnResultsParsingTypes()
{
  std::string list = "[";
  for (unsigned int i = 0; i < vpDetectorDNNOpenCV::COUNT - 1; i++) {
    list += "\"" + dnnResultsParsingTypeToString((vpDetectorDNNOpenCV::DNNResultsParsingType)i) + "\", ";
  }
  list += "\"" + dnnResultsParsingTypeToString((vpDetectorDNNOpenCV::DNNResultsParsingType)(vpDetectorDNNOpenCV::COUNT - 1)) + "\"]";
  return list;
}
 
std::string vpDetectorDNNOpenCV::dnnResultsParsingTypeToString(const DNNResultsParsingType &type)
{
  std::string name;
  switch (type) {
  case YOLO_V3:
    name = "yolov3";
    break;
  case YOLO_V4:
    name = "yolov4";
    break;
  case YOLO_V5:
    name = "yolov5";
    break;
  case YOLO_V7:
    name = "yolov7";
    break;
  case YOLO_V8:
    name = "yolov8";
    break;
  case YOLO_V11:
    name = "yolov11";
    break;
  case FASTER_RCNN:
    name = "faster-rcnn";
    break;
  case SSD_MOBILENET:
    name = "ssd-mobilenet";
    break;
  case RESNET_10:
    name = "resnet-10";
    break;
  case USER_SPECIFIED:
    name = "user-specified";
    break;
  case COUNT:
    name = "unknown";
    break;
  }
  return name;
}
 
vpDetectorDNNOpenCV::DNNResultsParsingType vpDetectorDNNOpenCV::dnnResultsParsingTypeFromString(const std::string &name)
{
  vpDetectorDNNOpenCV::DNNResultsParsingType res(COUNT);
  bool hasFoundMatch = false;
  std::string name_lowercase = vpIoTools::toLowerCase(name);
  for (int id = 0; id < COUNT && !hasFoundMatch; id++) {
    vpDetectorDNNOpenCV::DNNResultsParsingType temp = (vpDetectorDNNOpenCV::DNNResultsParsingType)id;
    if (dnnResultsParsingTypeToString(temp) == name_lowercase) {
      res = temp;
      hasFoundMatch = true;
    }
  }
  return res;
}
 
std::vector<std::string> vpDetectorDNNOpenCV::parseClassNamesFile(const std::string &filename)
{
  return NetConfig::parseClassNamesFile(filename);
}
 
vpDetectorDNNOpenCV::vpDetectorDNNOpenCV()
  : m_applySizeFilterAfterNMS(false), m_blob(), m_I_color(), m_img(),
  m_net(), m_netConfig(), m_outNames(), m_dnnRes(),
  m_parsingMethod(vpDetectorDNNOpenCV::postProcess_unimplemented)
{
  setDetectionFilterSizeRatio(m_netConfig.m_filterSizeRatio);
}
 
vpDetectorDNNOpenCV::vpDetectorDNNOpenCV(const NetConfig &config, const DNNResultsParsingType &typeParsingMethod, void (*parsingMethod)(DetectionCandidates &, std::vector<cv::Mat> &, const NetConfig &))
  : m_applySizeFilterAfterNMS(false), m_blob(), m_I_color(), m_img(),
  m_net(), m_netConfig(config), m_outNames(), m_dnnRes()
{
  setDetectionFilterSizeRatio(m_netConfig.m_filterSizeRatio);
  setParsingMethod(typeParsingMethod, parsingMethod);
  if (!m_netConfig.m_modelFilename.empty()) {
    readNet(m_netConfig.m_modelFilename, m_netConfig.m_modelConfigFilename, m_netConfig.m_framework);
  }
}
 
#ifdef VISP_HAVE_NLOHMANN_JSON
 
using json = nlohmann::json;
 
vpDetectorDNNOpenCV::vpDetectorDNNOpenCV(const std::string &jsonPath, void (*parsingMethod)(DetectionCandidates &, std::vector<cv::Mat> &, const NetConfig &))
  : m_applySizeFilterAfterNMS(false), m_blob(), m_I_color(), m_img(),
  m_net(), m_netConfig(), m_outNames(), m_dnnRes()
{
  initFromJSON(jsonPath);
  setDetectionFilterSizeRatio(m_netConfig.m_filterSizeRatio);
  setParsingMethod(m_netConfig.m_parsingMethodType, parsingMethod);
}
 
void vpDetectorDNNOpenCV::initFromJSON(const std::string &jsonPath)
{
  std::ifstream file(jsonPath);
  if (!file.good()) {
    std::stringstream ss;
    ss << "Problem opening file " << jsonPath << ". Make sure it exists and is readable" << std::endl;
    throw vpException(vpException::ioError, ss.str());
  }
  json j;
  try {
    j = json::parse(file);
  }
  catch (json::parse_error &e) {
    std::stringstream msg;
    msg << "Could not parse JSON file : \n";
 
    msg << e.what() << std::endl;
    msg << "Byte position of error: " << e.byte;
    throw vpException(vpException::ioError, msg.str());
  }
  *this = j; // Call from_json(const json& j, vpDetectorDNN& *this) to read json
  file.close();
  readNet(m_netConfig.m_modelFilename, m_netConfig.m_modelConfigFilename, m_netConfig.m_framework);
}
 
void vpDetectorDNNOpenCV::saveConfigurationInJSON(const std::string &jsonPath) const
{
  std::ofstream file(jsonPath);
  const json j = *this;
  file << j.dump(4);
  file.close();
}
#endif
 
vpDetectorDNNOpenCV::~vpDetectorDNNOpenCV() { }
 
bool vpDetectorDNNOpenCV::detect(const vpImage<unsigned char> &I, std::vector<DetectedFeatures2D> &output)
{
  vpImageConvert::convert(I, m_I_color);
 
  return detect(m_I_color, output);
}
 
bool vpDetectorDNNOpenCV::detect(const vpImage<unsigned char> &I, std::map< std::string, std::vector<DetectedFeatures2D>> &output)
{
  vpImageConvert::convert(I, m_I_color);
 
  return detect(m_I_color, output);
}
 
bool vpDetectorDNNOpenCV::detect(const vpImage<unsigned char> &I, std::vector< std::pair<std::string, std::vector<DetectedFeatures2D>>> &output)
{
  vpImageConvert::convert(I, m_I_color);
 
  return detect(m_I_color, output);
}
 
bool vpDetectorDNNOpenCV::detect(const vpImage<vpRGBa> &I, std::vector<DetectedFeatures2D> &output)
{
  vpImageConvert::convert(I, m_img);
 
  return detect(m_img, output);
}
 
bool vpDetectorDNNOpenCV::detect(const vpImage<vpRGBa> &I, std::map< std::string, std::vector<DetectedFeatures2D>> &output)
{
  vpImageConvert::convert(I, m_img);
 
  return detect(m_img, output);
}
 
bool vpDetectorDNNOpenCV::detect(const vpImage<vpRGBa> &I, std::vector< std::pair<std::string, std::vector<DetectedFeatures2D>>> &output)
{
  vpImageConvert::convert(I, m_img);
 
  return detect(m_img, output);
}
 
bool vpDetectorDNNOpenCV::detect(const cv::Mat &I, std::vector<DetectedFeatures2D> &output)
{
  m_img = I;
  output.clear();
 
  cv::Size inputSize(m_netConfig.m_inputSize.width > 0 ? m_netConfig.m_inputSize.width : m_img.cols,
    m_netConfig.m_inputSize.height > 0 ? m_netConfig.m_inputSize.height : m_img.rows);
  cv::dnn::blobFromImage(m_img, m_blob, m_netConfig.m_scaleFactor, inputSize, m_netConfig.m_mean, m_netConfig.m_swapRB, false);
 
  m_net.setInput(m_blob);
  try {
    m_net.forward(m_dnnRes, m_outNames);
  }
  catch (const cv::Exception &e) {
    std::cerr << "Caught an exception trying to run inference:" << std::endl << "\t"
      << e.what()
      << "\nCuda and/or GPU driver might not be correctly installed. Setting preferable backend to CPU and trying again." << std::endl;
    m_net.setPreferableBackend(cv::dnn::DNN_BACKEND_DEFAULT);
    m_net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
    m_net.forward(m_dnnRes, m_outNames);
  }
 
  DetectionCandidates proposals;
  postProcess(proposals);
  size_t nbClassNames = m_netConfig.m_classNames.size();
  for (size_t i = 0; i < m_indices.size(); ++i) {
    int idx = m_indices[i];
    cv::Rect box = proposals.m_boxes[idx];
    std::optional<std::string> classname_opt;
    if (nbClassNames > 0) {
      classname_opt = m_netConfig.m_classNames[proposals.m_classIds[idx]];
    }
    output.emplace_back(box.x, box.x + box.width, box.y, box.y + box.height
      , proposals.m_classIds[idx], proposals.m_confidences[idx]
      , classname_opt
    );
  }
 
  if (m_applySizeFilterAfterNMS) {
    // removing false detections, based on the bbox sizes
    output = filterDetectionMultiClassInput(output, m_netConfig.m_filterSizeRatio);
  }
 
  return !output.empty();
}
 
bool vpDetectorDNNOpenCV::detect(const cv::Mat &I, std::map< std::string, std::vector<DetectedFeatures2D>> &output)
{
  m_img = I;
  output.clear();
 
  cv::Size inputSize(m_netConfig.m_inputSize.width > 0 ? m_netConfig.m_inputSize.width : m_img.cols,
    m_netConfig.m_inputSize.height > 0 ? m_netConfig.m_inputSize.height : m_img.rows);
  cv::dnn::blobFromImage(m_img, m_blob, m_netConfig.m_scaleFactor, inputSize, m_netConfig.m_mean, m_netConfig.m_swapRB, false);
 
  m_net.setInput(m_blob);
  try {
    m_net.forward(m_dnnRes, m_outNames);
  }
  catch (const cv::Exception &e) {
    std::cerr << "Caught an exception trying to run inference:" << std::endl << "\t"
      << e.what()
      << "\nCuda and/or GPU driver might not be correctly installed. Setting preferable backend to CPU and trying again." << std::endl;
    m_net.setPreferableBackend(cv::dnn::DNN_BACKEND_DEFAULT);
    m_net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
    m_net.forward(m_dnnRes, m_outNames);
  }
 
  DetectionCandidates proposals;
  postProcess(proposals);
  size_t nbClassNames = m_netConfig.m_classNames.size();
  for (size_t i = 0; i < m_indices.size(); ++i) {
    int idx = m_indices[i];
    cv::Rect box = proposals.m_boxes[idx];
    std::string classname;
    if (nbClassNames > 0) {
      classname = m_netConfig.m_classNames[proposals.m_classIds[idx]];
    }
    else {
      classname = std::to_string(proposals.m_classIds[idx]);
    }
    std::optional<std::string> classname_opt = std::optional<std::string>(classname);
    output[classname].emplace_back(box.x, box.x + box.width, box.y, box.y + box.height
      , proposals.m_classIds[idx], proposals.m_confidences[idx]
      , classname_opt
    );
  }
 
  if (m_applySizeFilterAfterNMS) {
    output = filterDetectionMultiClassInput(output, m_netConfig.m_filterSizeRatio);
  }
 
  return !output.empty();
}
 
bool vpDetectorDNNOpenCV::detect(const cv::Mat &I, std::vector< std::pair<std::string, std::vector<DetectedFeatures2D>>> &output)
{
  std::map< std::string, std::vector<DetectedFeatures2D>> map_output;
  bool returnStatus = detect(I, map_output);
  for (auto key_val : map_output) {
    output.push_back(key_val);
  }
  return returnStatus;
}
 
#if (VISP_HAVE_OPENCV_VERSION == 0x030403)
std::vector<cv::String> vpDetectorDNNOpenCV::getOutputsNames()
{
  static std::vector<cv::String> names;
  if (names.empty()) {
    std::vector<int> outLayers = m_net.getUnconnectedOutLayers();
    std::vector<cv::String> layersNames = m_net.getLayerNames();
    names.resize(outLayers.size());
    for (size_t i = 0; i < outLayers.size(); ++i)
      names[i] = layersNames[outLayers[i] - 1];
  }
  return names;
}
#endif
 
void vpDetectorDNNOpenCV::postProcess(DetectionCandidates &proposals)
{
  switch (m_netConfig.m_parsingMethodType) {
  case YOLO_V3:
  case YOLO_V4:
    postProcess_YoloV3_V4(proposals, m_dnnRes, m_netConfig);
    break;
  case YOLO_V5:
  case YOLO_V7:
    postProcess_YoloV5_V7(proposals, m_dnnRes, m_netConfig);
    break;
  case YOLO_V8:
  case YOLO_V11:
    postProcess_YoloV8_V11(proposals, m_dnnRes, m_netConfig);
    break;
  case FASTER_RCNN:
    postProcess_FasterRCNN(proposals, m_dnnRes, m_netConfig);
    break;
  case SSD_MOBILENET:
#if defined(VISP_BUILD_DEPRECATED_FUNCTIONS)
    void postProcess_SSD_MobileNet(DetectionCandidates & proposals, std::vector<cv::Mat> &dnnRes, const NetConfig & netConfig);
#else
    // NB: the two SSD-MobileNet DNNs that have been tested worked only
    // using the ResNet-10 parsing method
    postProcess_ResNet_10(proposals, m_dnnRes, m_netConfig);
#endif
    break;
  case RESNET_10:
    postProcess_ResNet_10(proposals, m_dnnRes, m_netConfig);
    break;
  case USER_SPECIFIED:
    m_parsingMethod(proposals, m_dnnRes, m_netConfig);
    break;
  default:
    throw(vpException(vpException::badValue, "Type of DNN post-processing method not handled."));
  }
 
  m_indices.clear();
  cv::dnn::NMSBoxes(proposals.m_boxes, proposals.m_confidences, m_netConfig.m_confThreshold, m_netConfig.m_nmsThreshold, m_indices);
}
 
std::vector<vpDetectorDNNOpenCV::DetectedFeatures2D>
vpDetectorDNNOpenCV::filterDetectionSingleClassInput(const std::vector<DetectedFeatures2D> &detected_features, const double minRatioOfAreaOk)
{
  double meanArea(0.);
  double originalNumberOfObj = static_cast<double>(detected_features.size());
  double meanFactor = 1. / originalNumberOfObj;
 
  // Computing the average area of the class
  for (DetectedFeatures2D feature : detected_features) {
    meanArea += feature.m_bbox.getArea();
  }
  meanArea *= meanFactor;
 
  // Keeping only the detections that respect the area criterion
  std::vector<DetectedFeatures2D> filtered_features;
  for (DetectedFeatures2D feature : detected_features) {
    if (feature.m_bbox.getArea() >= minRatioOfAreaOk * meanArea && feature.m_bbox.getArea() < meanArea / minRatioOfAreaOk) {
      filtered_features.push_back(feature);
    }
  }
 
  return filtered_features;
}
 
std::vector<vpDetectorDNNOpenCV::DetectedFeatures2D>
vpDetectorDNNOpenCV::filterDetectionMultiClassInput(const std::vector<DetectedFeatures2D> &detected_features, const double minRatioOfAreaOk)
{
#ifndef DOXYGEN_SHOULD_SKIP_THIS
  class MeanAreaComputer
  {
  private:
    std::map<int, std::pair<int, double>> m_map_id_pairOccurrencesAreas; 
    std::map<int, double> m_mapMeans; 
    double computeMeanArea(const int &class_id)
    {
      return m_map_id_pairOccurrencesAreas[class_id].second / (double)m_map_id_pairOccurrencesAreas[class_id].first;
    }
 
  public:
    void computeMeans()
    {
      for (const auto &classID_pair : m_map_id_pairOccurrencesAreas) {
        m_mapMeans[classID_pair.first] = computeMeanArea(classID_pair.first);
      }
    }
 
    double getMean(const int &class_id)
    {
      if (m_map_id_pairOccurrencesAreas.find(class_id) == m_map_id_pairOccurrencesAreas.end()) {
        throw(vpException(vpException::badValue, "[MeanAreaComputer::getMean] Asking for class_id \"" + std::to_string(class_id) + "\" that is not present in m_mapMeans. Did you call computeMeans ?"));
      }
      return m_mapMeans[class_id];
    }
 
    void operator()(const DetectedFeatures2D &feature)
    {
      int class_id = feature.getClassId();
      double area = feature.getBoundingBox().getArea();
      if (m_map_id_pairOccurrencesAreas.find(class_id) == m_map_id_pairOccurrencesAreas.end()) {
        m_map_id_pairOccurrencesAreas[class_id] = std::pair<int, double>(1, area);
      }
      else {
        std::pair<int, double> prev_state = m_map_id_pairOccurrencesAreas[class_id];
        m_map_id_pairOccurrencesAreas[class_id] = std::pair<int, double>(prev_state.first + 1, prev_state.second + area);
      }
    }
  };
#endif // DOXYGEN_SHOULD_SKIP_THIS
 
  // Computing the average area of each class
  MeanAreaComputer meanComputer;
  std::for_each(detected_features.begin(), detected_features.end(), meanComputer);
  meanComputer.computeMeans();
 
  // Keeping only the detections that respect the area criterion
  std::vector<DetectedFeatures2D> filtered_features;
  for (DetectedFeatures2D feature : detected_features) {
    double meanArea = meanComputer.getMean(feature.getClassId());
    if (feature.m_bbox.getArea() >= minRatioOfAreaOk * meanArea
      && feature.m_bbox.getArea() < meanArea / minRatioOfAreaOk) {
      filtered_features.push_back(feature);
    }
  }
 
  return filtered_features;
}
 
std::map<std::string, std::vector<vpDetectorDNNOpenCV::DetectedFeatures2D>>
vpDetectorDNNOpenCV::filterDetectionMultiClassInput(const std::map< std::string, std::vector<vpDetectorDNNOpenCV::DetectedFeatures2D>> &detected_features, const double minRatioOfAreaOk)
{
  std::map<std::string, std::vector<vpDetectorDNNOpenCV::DetectedFeatures2D>> output;
  for (auto keyval : detected_features) {
    output[keyval.first] = filterDetectionSingleClassInput(detected_features.at(keyval.first), minRatioOfAreaOk); // removing false detections
  }
  return output;
}
 
void vpDetectorDNNOpenCV::postProcess_YoloV3_V4(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  size_t nbBatches = dnnRes.size();
 
  for (size_t i = 0; i < nbBatches; i++) {
    // Slightly modify from here: https://github.com/opencv/opencv/blob/8c25a8eb7b10fb50cda323ee6bec68aa1a9ce43c/samples/dnn/object_detection.cpp#L192-L221
     // Counts the number of proposed detections and the number of data corresponding to 1 detection
    int num_proposal = dnnRes[i].size[0]; // Number of detections
    int nout = dnnRes[i].size[1]; // Number of data for each detection
    if (dnnRes[i].dims > 2) {
      num_proposal = dnnRes[i].size[1];
      nout = dnnRes[i].size[2];
      dnnRes[i] = dnnRes[i].reshape(0, num_proposal);
    }
 
    int n = 0, row_ind = 0; 
    float *pdata = (float *)dnnRes[i].data;
 
    // Iterate on the detections to keep only the meaningful ones
    for (n = 0; n < num_proposal; n++) {
      float box_score = pdata[4];
      if (box_score > netConfig.m_confThreshold) {
        cv::Mat scores = dnnRes[i].row(row_ind).colRange(5, nout);
        cv::Point classIdPoint;
        double max_class_score;
        // Get the value and location of the maximum score
        cv::minMaxLoc(scores, 0, &max_class_score, 0, &classIdPoint);
 
        max_class_score *= box_score;
 
        // The detection is kept only if the confidence is greater than the threshold
        if (max_class_score > netConfig.m_confThreshold) {
          const int class_idx = classIdPoint.x;
          float cx = pdata[0] * m_img.cols; 
          float cy = pdata[1] * m_img.rows; 
          float w = pdata[2] * m_img.cols; 
          float h = pdata[3] * m_img.rows; 
 
          int left = int(cx - 0.5 * w);
          int top = int(cy - 0.5 * h);
 
          proposals.m_confidences.push_back((float)max_class_score);
          proposals.m_boxes.push_back(cv::Rect(left, top, (int)(w), (int)(h)));
          proposals.m_classIds.push_back(class_idx);
        }
      }
      row_ind++;
      pdata += nout;
    }
  }
}
 
void vpDetectorDNNOpenCV::postProcess_YoloV5_V7(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  // Compute the ratio between the original size of the image and the network size to translate network coordinates into
  // image coordinates
  float ratioh = (float)m_img.rows / netConfig.m_inputSize.height, ratiow = (float)m_img.cols / netConfig.m_inputSize.width;
  size_t nbBatches = dnnRes.size();
 
  for (size_t i = 0; i < nbBatches; i++) {
    // Counts the number of proposed detections and the number of data corresponding to 1 detection
    int num_proposal = dnnRes[i].size[0]; // Number of detections
    int nout = dnnRes[i].size[1]; // Number of data for each detection
    if (dnnRes[i].dims > 2) {
      num_proposal = dnnRes[i].size[1];
      nout = dnnRes[i].size[2];
      dnnRes[i] = dnnRes[i].reshape(0, num_proposal);
    }
 
    int n = 0, row_ind = 0; 
    float *pdata = (float *)dnnRes[i].data;
 
    // Iterate on the detections to keep only the meaningful ones
    for (n = 0; n < num_proposal; n++) {
      float box_score = pdata[4];
 
      if (box_score > netConfig.m_confThreshold) {
        cv::Mat scores = dnnRes[i].row(row_ind).colRange(5, nout);
        cv::Point classIdPoint;
        double max_class_score;
        // Get the value and location of the maximum score
        cv::minMaxLoc(scores, 0, &max_class_score, 0, &classIdPoint);
        max_class_score *= box_score;
 
        // The detection is kept only if the confidence is greater than the threshold
        if (max_class_score > netConfig.m_confThreshold) {
          const int class_idx = classIdPoint.x;
          float cx = pdata[0] * ratiow; 
          float cy = pdata[1] * ratioh; 
          float w = pdata[2] * ratiow; 
          float h = pdata[3] * ratioh; 
 
          int left = int(cx - 0.5 * w);
          int top = int(cy - 0.5 * h);
 
          proposals.m_confidences.push_back((float)max_class_score);
          proposals.m_boxes.push_back(cv::Rect(left, top, (int)(w), (int)(h)));
          proposals.m_classIds.push_back(class_idx);
        }
      }
      row_ind++;
      pdata += nout;
    }
  }
}
 
void vpDetectorDNNOpenCV::postProcess_YoloV8_V11(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  // Code adapted from here: https://github.com/JustasBart/yolov8_CPP_Inference_OpenCV_ONNX/blob/minimalistic/inference.cpp
  // Compute the ratio between the original size of the image and the network size to translate network coordinates into
  // image coordinates
  float ratioh = (float)m_img.rows / netConfig.m_inputSize.height, ratiow = (float)m_img.cols / netConfig.m_inputSize.width;
  size_t nbBatches = dnnRes.size();
 
  for (size_t i = 0; i < nbBatches; i++) {
    // Counts the number of proposed detections and the number of data corresponding to 1 detection
    int num_proposal = dnnRes[i].size[1]; // Number of detections
    int nout = dnnRes[i].size[0]; // Number of data for each detection
    if (dnnRes[i].dims > 2) {
      num_proposal = dnnRes[i].size[2];
      nout = dnnRes[i].size[1];
      dnnRes[i] = dnnRes[i].reshape(0, nout);
    }
    cv::transpose(dnnRes[i], dnnRes[i]); // Organise data as YoloV5 i.e. [batchsize][1:num_proposals][1:4+nb_classes]
 
    int n = 0, row_ind = 0; 
    float *pdata = (float *)dnnRes[i].data;
 
    // Iterate on the detections to keep only the meaningful ones
    for (n = 0; n < num_proposal; n++) {
      cv::Mat scores = dnnRes[i].row(row_ind).colRange(4, nout);
      cv::Point classIdPoint;
      double max_class_score;
      // Get the value and location of the maximum score
      cv::minMaxLoc(scores, 0, &max_class_score, 0, &classIdPoint);
 
      // The detection is kept only if the confidence is greater than the threshold
      if (max_class_score > netConfig.m_confThreshold) {
        const int class_idx = classIdPoint.x;
        float cx = pdata[0] * ratiow; 
        float cy = pdata[1] * ratioh; 
        float w = pdata[2] * ratiow; 
        float h = pdata[3] * ratioh; 
 
        int left = int(cx - 0.5 * w);
        int top = int(cy - 0.5 * h);
 
        proposals.m_confidences.push_back((float)max_class_score);
        proposals.m_boxes.push_back(cv::Rect(left, top, (int)(w), (int)(h)));
        proposals.m_classIds.push_back(class_idx);
      }
 
      row_ind++;
      pdata += nout;
    }
  }
}
 
void vpDetectorDNNOpenCV::postProcess_FasterRCNN(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  // Direct copy from object_detection.cpp OpenCV sample
  // Faster-RCNN
 
  // 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]
  size_t nbBatches = dnnRes.size();
  for (size_t j = 0; j < nbBatches; j++) {
    float *data = (float *)dnnRes[j].data;
    for (size_t i = 0; i < dnnRes[j].total(); i += 7) {
      float confidence = data[i + 2];
      if (confidence > netConfig.m_confThreshold) {
        int left = (int)(data[i + 3] * m_img.cols);
        int top = (int)(data[i + 4] * m_img.rows);
        int right = (int)(data[i + 5] * m_img.cols);
        int bottom = (int)(data[i + 6] * m_img.rows);
        int classId = (int)(data[i + 1]);
 
        proposals.m_confidences.push_back((float)confidence);
        proposals.m_boxes.push_back(cv::Rect(left, top, right - left + 1, bottom - top + 1));
        proposals.m_classIds.push_back(classId);
      }
    }
  }
 
}
 
#if defined(VISP_BUILD_DEPRECATED_FUNCTIONS)
void vpDetectorDNNOpenCV::postProcess_SSD_MobileNet(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  // Network produces 2 outputs blob:
  // - `scores` with dimensions 1xNxC
  // - 'boxes'  with dimensions 1xNx4
  // where `N` is a number of detections and `C` is the number of classes (with `BACKGROUND` as classId = 0).
 
  int scores_index = m_outNames[0] == "scores" ? 0 : 1; // scores output index.
  int boxes_index = m_outNames[0] == "boxes" ? 0 : 1;   // boxes output index.
 
  int N = dnnRes[scores_index].size[1], C = dnnRes[scores_index].size[2];
 
  float *confidence = (float *)dnnRes[scores_index].data;
  float *bbox = (float *)dnnRes[boxes_index].data;
 
  // Loop over all guesses on the output of the network.
  for (int i = 0; i < N; i++) {
    uint32_t maxClass = 0;
    float maxScore = -1000.0f;
 
    for (int j = 1; j < C; j++) // ignore background (classId = 0).
    {
      const float score = confidence[i * C + j];
 
      if (score < netConfig.m_confThreshold)
        continue;
 
      if (score > maxScore) {
        maxScore = score;
        maxClass = j;
      }
    }
 
    if (maxScore > netConfig.m_confThreshold) {
      int left = (int)(bbox[4 * i] * m_img.cols);
      int top = (int)(bbox[4 * i + 1] * m_img.rows);
      int right = (int)(bbox[4 * i + 2] * m_img.cols);
      int bottom = (int)(bbox[4 * i + 3] * m_img.rows);
      int width = right - left + 1;
      int height = bottom - top + 1;
 
      int classId = maxClass;
      proposals.m_confidences.push_back(maxScore);
      proposals.m_boxes.push_back(cv::Rect(left, top, width, height));
      proposals.m_classIds.push_back(classId);
    }
  }
}
#endif
 
void vpDetectorDNNOpenCV::postProcess_ResNet_10(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  // Direct copy from object_detection.cpp OpenCV sample
 
  // 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(dnnRes.size() == 1);
  float *data = (float *)dnnRes[0].data;
  for (size_t i = 0; i < dnnRes[0].total(); i += 7) {
    float confidence = data[i + 2];
    if (confidence > netConfig.m_confThreshold) {
      int left = (int)(data[i + 3] * m_img.cols);
      int top = (int)(data[i + 4] * m_img.rows);
      int right = (int)(data[i + 5] * m_img.cols);
      int bottom = (int)(data[i + 6] * m_img.rows);
      int classId = (int)(data[i + 1]) - 1;
 
      proposals.m_confidences.push_back((float)confidence);
      proposals.m_boxes.push_back(cv::Rect(left, top, right - left + 1, bottom - top + 1));
      proposals.m_classIds.push_back(classId);
    }
  }
}
 
void vpDetectorDNNOpenCV::postProcess_unimplemented(DetectionCandidates &proposals, std::vector<cv::Mat> &dnnRes, const NetConfig &netConfig)
{
  (void)proposals;
  (void)dnnRes;
  (void)netConfig;
  throw(vpException(vpException::functionNotImplementedError, "vpDetectorDNNOpenCV::postProcess was called with a USER_SPECIFIED DNN but not post processing method was given."));
}
 
void vpDetectorDNNOpenCV::readNet(const std::string &model, const std::string &config, const std::string &framework)
{
  m_netConfig.m_modelFilename = model;
  m_netConfig.m_modelConfigFilename = config;
  m_netConfig.m_framework = framework;
  m_net = cv::dnn::readNet(model, config, framework);
#if (VISP_HAVE_OPENCV_VERSION == 0x030403)
  m_outNames = getOutputsNames();
#else
  m_outNames = m_net.getUnconnectedOutLayersNames();
#endif
}
 
void vpDetectorDNNOpenCV::setNetConfig(const NetConfig &config)
{
  m_netConfig = config;
  setDetectionFilterSizeRatio(m_netConfig.m_filterSizeRatio);
  setParsingMethod(m_netConfig.m_parsingMethodType);
  if (!m_netConfig.m_modelFilename.empty()) {
    readNet(m_netConfig.m_modelFilename, m_netConfig.m_modelConfigFilename, m_netConfig.m_framework);
  }
}
 
void vpDetectorDNNOpenCV::setConfidenceThreshold(const float &confThreshold) { m_netConfig.m_confThreshold = confThreshold; }
 
void vpDetectorDNNOpenCV::setNMSThreshold(const float &nmsThreshold) { m_netConfig.m_nmsThreshold = nmsThreshold; }
 
void vpDetectorDNNOpenCV::setDetectionFilterSizeRatio(const double &sizeRatio)
{
  m_netConfig.m_filterSizeRatio = sizeRatio;
  if (m_netConfig.m_filterSizeRatio > std::numeric_limits<double>::epsilon()) {
    m_applySizeFilterAfterNMS = true;
  }
  else {
    m_applySizeFilterAfterNMS = false;
  }
}
 
void vpDetectorDNNOpenCV::setInputSize(const int &width, const int &height)
{
  m_netConfig.m_inputSize.width = width;
  m_netConfig.m_inputSize.height = height;
}
 
void vpDetectorDNNOpenCV::setMean(const double &meanR, const double &meanG, const double &meanB) { m_netConfig.m_mean = cv::Scalar(meanR, meanG, meanB); }
 
void vpDetectorDNNOpenCV::setPreferableBackend(const int &backendId) { m_net.setPreferableBackend(backendId); }
 
void vpDetectorDNNOpenCV::setPreferableTarget(const int &targetId) { m_net.setPreferableTarget(targetId); }
 
void vpDetectorDNNOpenCV::setScaleFactor(const double &scaleFactor)
{
  m_netConfig.m_scaleFactor = scaleFactor;
  if ((m_netConfig.m_parsingMethodType == YOLO_V7 || m_netConfig.m_parsingMethodType == YOLO_V8 || m_netConfig.m_parsingMethodType == YOLO_V11) && m_netConfig.m_scaleFactor != 1 / 255.) {
    std::cout << "[vpDetectorDNNOpenCV::setParsingMethod] WARNING: scale factor should be 1/255. to normalize pixels value." << std::endl;
  }
}
 
void vpDetectorDNNOpenCV::setSwapRB(const bool &swapRB) { m_netConfig.m_swapRB = swapRB; }
 
void vpDetectorDNNOpenCV::setParsingMethod(const DNNResultsParsingType &typeParsingMethod, void (*parsingMethod)(DetectionCandidates &, std::vector<cv::Mat> &, const NetConfig &))
{
  m_netConfig.m_parsingMethodType = typeParsingMethod;
  m_parsingMethod = parsingMethod;
  if ((m_netConfig.m_parsingMethodType == YOLO_V7 || m_netConfig.m_parsingMethodType == YOLO_V8 || m_netConfig.m_parsingMethodType == YOLO_V11) && m_netConfig.m_scaleFactor != 1 / 255.) {
    m_netConfig.m_scaleFactor = 1 / 255.;
    std::cout << "[vpDetectorDNNOpenCV::setParsingMethod] NB: scale factor changed to 1/255. to normalize pixels value." << std::endl;
  }
 
#if defined(VISP_BUILD_DEPRECATED_FUNCTIONS)
  if (m_netConfig.m_parsingMethodType == SSD_MOBILENET) {
    std::cout << "[vpDetectorDNNOpenCV::setParsingMethod] WARNING: The chosen type of network is " << dnnResultsParsingTypeToString(m_netConfig.m_parsingMethodType) << " VISP_BUILD_DEPRECATED_FUNCTIONS is set to true." << std::endl;
    std::cout << "\tThe parsing method that worked with  the networks quoted in the ViSP documentation was postProcess_ResNet_10 instead of postProcess_SSD_MobileNet." << std::endl;
    std::cout << "\tIf the SSD-MobileNet network does not seem to work, please try to recompile ViSP setting VISP_BUILD_DEPRECATED_FUNCTIONS as false." << std::endl << std::flush;
  }
#endif
}
 
END_VISP_NAMESPACE
#elif !defined(VISP_BUILD_SHARED_LIBS)
// Work around to avoid warning: libvisp_core.a(vpDetectorDNNOpenCV.cpp.o) has no symbols
void dummy_vpDetectorDNN() { };
#endif