TorchTagDetection.hpp

 
#ifndef TORCH_TAG_DETECTION_HPP
#define TORCH_TAG_DETECTION_HPP
 
#include "../../util/vision/TagDetectionUtilty.hpp"
#include "../../util/vision/YOLOModel.hpp"
 
#include <chrono>
 
 
 
namespace torchtag
{
 
    inline std::vector<tagdetectutils::ArucoTag> Detect(const cv::Mat& cvFrame,
                                                        yolomodel::pytorch::PyTorchInterpreter& trPyTorchDetector,
                                                        const float fMinObjectConfidence = 0.40f,
                                                        const float fNMSThreshold        = 0.60f)
    {
        // Check if the input frame is in RGB format.
        if (cvFrame.channels() != 3)
        {
            // Submit logger message.
            LOG_ERROR(logging::g_qSharedLogger, "Detect() requires a RGB image.");
            return {};
        }
 
        // Declare instance variables.
        std::vector<tagdetectutils::ArucoTag> vDetectedTags;
 
        // Check if the PyTorch interpreter hardware is opened and the model is loaded.
        if (trPyTorchDetector.IsReadyForInference())
        {
            // Run inference on YOLO model with current image.
            std::vector<yolomodel::Detection> vOutputTensorTags = trPyTorchDetector.Inference(cvFrame, fMinObjectConfidence, fNMSThreshold);
 
            // Repackage detections into tensorflow tags.
            for (const yolomodel::Detection& stTagDetection : vOutputTensorTags)
            {
                // Create and initialize new TensorflowTag.
                tagdetectutils::ArucoTag stDetectedTag;
                stDetectedTag.dConfidence       = stTagDetection.fConfidence;
                stDetectedTag.pBoundingBox      = std::make_shared<cv::Rect2d>(stTagDetection.cvBoundingBox);
                stDetectedTag.nID               = stTagDetection.nClassID;
                stDetectedTag.szClassName       = stTagDetection.szClassName;
                stDetectedTag.eDetectionMethod  = tagdetectutils::TagDetectionMethod::eTorch;
                stDetectedTag.cvImageResolution = cvFrame.size();
 
                // Add the newly detected tag to the vector.
                vDetectedTags.emplace_back(stDetectedTag);
            }
        }
        else
        {
            // Submit logger message.
            LOG_WARNING(logging::g_qSharedLogger,
                        "TorchDetect: Unable to detect tags using YOLO torch detection because hardware is not opened or model is not initialized.");
        }
 
        // Return the detected tags.
        return vDetectedTags;
    }
 
 
    inline void DrawDetections(cv::Mat& cvDetectionsFrame, const std::vector<tagdetectutils::ArucoTag>& vDetectedTags)
    {
        // Check if the given frame is a 1 or 3 channel image. (not BGRA)
        if (!cvDetectionsFrame.empty() && (cvDetectionsFrame.channels() == 1 || cvDetectionsFrame.channels() == 3))
        {
            // Loop through each detection.
            for (const tagdetectutils::ArucoTag& stTag : vDetectedTags)
            {
                // Check if the tag detection type is Torch.
                if (stTag.eDetectionMethod == tagdetectutils::TagDetectionMethod::eTorch)
                {
                    // Draw bounding box onto image.
                    cv::rectangle(cvDetectionsFrame, *stTag.pBoundingBox, cv::Scalar(255, 255, 255), 2);
                    std::string szText  = stTag.szClassName + " " + std::to_string(static_cast<int>(stTag.dConfidence * 100)) + "%";
                    cv::Size cvTextSize = cv::getTextSize(szText, cv::FONT_HERSHEY_SIMPLEX, 0.75, 1, nullptr);
                    // Draw classID background box onto image.
                    cv::rectangle(cvDetectionsFrame,
                                  stTag.pBoundingBox->tl() + cv::Point2d(0, stTag.pBoundingBox->height),
                                  stTag.pBoundingBox->tl() + cv::Point2d(cvTextSize.width, stTag.pBoundingBox->height + cvTextSize.height),
                                  cv::Scalar(255, 255, 255),
                                  cv::FILLED);
                    // Draw class text onto image.
                    cv::putText(cvDetectionsFrame,
                                szText,
                                stTag.pBoundingBox->tl() + cv::Point2d(0, stTag.pBoundingBox->height + cvTextSize.height),
                                cv::FONT_HERSHEY_SIMPLEX,
                                0.5,
                                cv::Scalar(0, 0, 0));
                }
            }
        }
        else
        {
            // Submit logger message.
            LOG_ERROR(logging::g_qSharedLogger,
                      "TorchDetect: Unable to draw markers on image because it is empty or because it has {} channels. (Should be 1 or 3)",
                      cvDetectionsFrame.channels());
        }
    }
}    // namespace torchtag
 
#endif