Build a Simple AI Object Detection with Qt

project(${APP_TARGET} LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

find_package(Qt6 6.8 REQUIRED COMPONENTS Quick Multimedia)

qt_standard_project_setup()

# ---- OpenCV Android SDK ----
if(ANDROID)
  set(OPENCV_ANDROID_SDK "${CMAKE_CURRENT_LIST_DIR}/third_party/Opencv-android-sdk")
  set(OpenCV_DIR "${OPENCV_ANDROID_SDK}/sdk/native/jni")
endif()

find_package(OpenCV REQUIRED)

# ---- App ----
qt_add_executable(${APP_TARGET}
  MANUAL_FINALIZATION
  main.cpp
)

qt_add_qml_module(${APP_TARGET}
  URI ${QML_URI}
  VERSION 1.0
  QML_FILES
      qml/Main.qml

  RESOURCES
  NO_RESOURCE_TARGET_PATH
  SOURCES
    FrameGrabber.h FrameGrabber.cpp
    DetectorWorker.h DetectorWorker.cpp
    AnnotatedImageProvider.h AnnotatedImageProvider.cpp
    android/AndroidManifest.xml
)

# Embed model + labels into Qt resources at ":/assets/..."
qt_add_resources(${APP_TARGET} app_assets
  PREFIX "/assets"
  BASE "assets"
  FILES
    assets/model.onnx
    assets/labels.txt
)

target_link_libraries(${APP_TARGET}
  PRIVATE Qt6::Quick Qt6::Multimedia
  PRIVATE ${OpenCV_LIBS}
)

target_include_directories(${APP_TARGET}
  PRIVATE ${OpenCV_INCLUDE_DIRS}
)

if(ANDROID)
  set_property(TARGET ${APP_TARGET} PROPERTY
    QT_ANDROID_PACKAGE_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/android"
  )

  set(_opencv_so "${OPENCV_ANDROID_SDK}/sdk/native/libs/${ANDROID_ABI}/libopencv_java4.so")
  if(NOT EXISTS "${_opencv_so}")
    set(_opencv_so "${OPENCV_ANDROID_SDK}/sdk/native/libs/${ANDROID_ABI}/libopencv_java3.so")
  endif()

  if(NOT EXISTS "${_opencv_so}")
    message(FATAL_ERROR "OpenCV .so not found for ABI='${ANDROID_ABI}'")
  endif()

  set_property(TARGET ${APP_TARGET} APPEND PROPERTY
    QT_ANDROID_EXTRA_LIBS "${_opencv_so}"
  )

  target_link_libraries(${APP_TARGET} PRIVATE log)
endif()

if(ANDROID)
    set_target_properties(${APP_TARGET} PROPERTIES
            QT_ANDROID_PACKAGE_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/android"
            QT_ANDROID_PACKAGE_NAME "org.qtproject.QtAIDetectAndroid"
            QT_ANDROID_APP_NAME ${APP_TARGET}
            QT_ANDROID_TARGET_SDK_VERSION 31
            QT_ANDROID_MIN_SDK_VERSION 31
            QT_ANDROID_VERSION_NAME "1.0")

    qt_finalize_executable(${APP_TARGET})
endif()

install(TARGETS ${APP_TARGET}
    BUNDLE DESTINATION .
    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
)

qt_generate_deploy_qml_app_script(
    TARGET ${APP_TARGET}
    OUTPUT_SCRIPT deploy_script
    MACOS_BUNDLE_POST_BUILD
    NO_UNSUPPORTED_PLATFORM_ERROR
    DEPLOY_USER_QML_MODULES_ON_UNSUPPORTED_PLATFORM
)

install(SCRIPT ${deploy_script})

#pragma once
#include <QObject>
#include <QVideoSink>
#include <QImage>

class FrameGrabber : public QObject {
    Q_OBJECT
    Q_PROPERTY(int frameCounter READ frameCounter NOTIFY frameCounterChanged)

public:
    explicit FrameGrabber(QObject *parent = nullptr);
    QVideoSink* videoSink() { return &m_sink; }
    int frameCounter() const { return m_frameCounter; }

signals:
    void frameReady(const QImage &img);
    void frameCounterChanged();

public slots:
    void onAnnotatedFrameReady(const QImage &img);

private slots:
    void onVideoFrameChanged(const QVideoFrame &frame);

private:
    QVideoSink m_sink;
    int m_frameCounter = 0;

    // Simple throttle (optional): process only every Nth frame
    int m_skip = 1;     // set to 2 or 3 on slower phones
    int m_count = 0;
};

#include "FrameGrabber.h"
#include <QVideoFrame>

FrameGrabber::FrameGrabber(QObject *parent) : QObject(parent) {
    connect(&m_sink, &QVideoSink::videoFrameChanged,
            this, &FrameGrabber::onVideoFrameChanged);
}

void FrameGrabber::onVideoFrameChanged(const QVideoFrame &frame) {
    if (!frame.isValid())
        return;

    if (++m_count % m_skip != 0)
        return;

    QImage img = frame.toImage();

    if (img.isNull())
        return;

    emit frameReady(img);
}

void FrameGrabber::onAnnotatedFrameReady(const QImage &) {
    ++m_frameCounter;
    emit frameCounterChanged();
}

Step 4 — Object detection worker (OpenCV DNN + ONNX)

This worker:

• converts QImage → cv::Mat
• runs model
• applies threshold + NMS
• draws boxes and emits annotated image

#pragma once
#include <QObject>
#include <QImage>
#include <QMutex>
#include <QStringList>
#include <QVector>
#include <QRectF>
#include <opencv2/dnn.hpp>

class DetectorWorker : public QObject
{
    Q_OBJECT
public:
    explicit DetectorWorker(QObject *parent = nullptr);

public slots:
    void loadModel(const QString &onnxPath);

    // Optional configuration
    void loadLabels(const QString &labelsPath);          // supports ":/..." and "qrc:/..."
    void setInputSize(int w, int h);                     // e.g. 320,320
    void setUseLetterbox(bool enabled);                  // recommended for YOLO exports
    void setThresholds(float confThres, float nmsThres); // tune

    // Call for every camera frame. Internally runs one inference at a time
    // and keeps only the latest pending frame (latest-frame-wins).
    void processFrame(const QImage &img);

signals:
    void annotatedFrameReady(const QImage &img);
    void detectionsTextReady(const QStringList &lines);

private slots:
    void doWork(); // runs in worker thread

private:
    struct Det {
        QRectF box;
        int classId;
        float score;
    };

    struct LetterboxInfo {
        float r = 1.0f;  // scale
        int dw = 0;      // padding left
        int dh = 0;      // padding top
    };

    static cv::Mat qimageToBgrMat(const QImage &img);
    cv::Mat letterbox(const cv::Mat &srcBgr, LetterboxInfo &info) const;

    QVector<Det> runYoloFromOutput(const cv::Mat &outRaw, int imgW, int imgH, const LetterboxInfo &lb);
    QImage drawDetections(const QImage &src, const QVector<Det> &dets);

private:
    cv::dnn::Net m_net;
    bool m_ready = false;

    // Model / postprocess settings
    int m_inW = 320;
    int m_inH = 320;
    bool m_useLetterbox = true;

    float m_confThres = 0.20f; // lower default for 320
    float m_nmsThres  = 0.45f;

    QStringList m_labels;

    // Backpressure state
    bool  m_busy = false;
    QImage m_current;
    QImage m_pending;
    QMutex m_mutex;
};

#include "DetectorWorker.h"
#include <QDebug>
#include <QFile>
#include <QPainter>
#include <QMetaObject>
#include <opencv2/imgproc.hpp>
#include <opencv2/core.hpp>

static QString normalizeQrc(const QString &path)
{
    // QML often uses qrc:/..., QFile uses :/...
    if (path.startsWith("qrc:/"))
        return ":" + path.mid(3); // "qrc:/x" -> ":/x"
    return path;
}

static inline float sigmoidf(float x)
{
    if (x >= 0.0f) {
        const float z = std::exp(-x);
        return 1.0f / (1.0f + z);
    } else {
        const float z = std::exp(x);
        return z / (1.0f + z);
    }
}

static inline float scoreToProb(float s)
{
    // If the model outputs logits (<0 or >1), convert via sigmoid.
    if (s < 0.0f || s > 1.0f)
        return sigmoidf(s);
    return s;
}

cv::Mat DetectorWorker::qimageToBgrMat(const QImage &img)
{
    QImage tmp = img.convertToFormat(QImage::Format_RGBA8888);

    cv::Mat rgba(tmp.height(), tmp.width(), CV_8UC4,
                 const_cast<uchar*>(tmp.bits()),
                 tmp.bytesPerLine());

    cv::Mat bgr;
    cv::cvtColor(rgba, bgr, cv::COLOR_RGBA2BGR); // allocates bgr

    return bgr;
}

DetectorWorker::DetectorWorker(QObject *parent) : QObject(parent) {}

void DetectorWorker::setInputSize(int w, int h)
{
    if (w > 0) m_inW = w;
    if (h > 0) m_inH = h;
}

void DetectorWorker::setUseLetterbox(bool enabled)
{
    m_useLetterbox = enabled;
}

void DetectorWorker::setThresholds(float confThres, float nmsThres)
{
    m_confThres = confThres;
    m_nmsThres = nmsThres;
}

void DetectorWorker::loadLabels(const QString &labelsPath)
{
    const QString p = normalizeQrc(labelsPath);
    QFile f(p);

    if (!f.open(QIODevice::ReadOnly | QIODevice::Text)) {
        qWarning() << "Failed to open labels file:" << labelsPath << f.errorString();

        return;
    }

    QStringList labels;

    while (!f.atEnd()) {
        const QString line = QString::fromUtf8(f.readLine()).trimmed();
        if (!line.isEmpty())
            labels << line;
    }

    f.close();

    m_labels = labels;
    qDebug() << "Loaded labels:" << m_labels.size() << "from" << labelsPath;
}

void DetectorWorker::loadModel(const QString &onnxPath)
{
    m_ready = false;
    try {
        m_net = cv::dnn::readNetFromONNX(onnxPath.toStdString());
        m_net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        m_net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
        m_ready = true;

        qDebug() << "Model loaded:" << onnxPath;
    } catch (const cv::Exception &e) {
        qWarning() << "Model load failed:" << e.what();
        m_ready = false;
    }
}

cv::Mat DetectorWorker::letterbox(const cv::Mat &srcBgr, LetterboxInfo &info) const
{
    const int srcW = srcBgr.cols;
    const int srcH = srcBgr.rows;
    const float r = std::min((float)m_inW / (float)srcW, (float)m_inH / (float)srcH);
    const int newW = (int)std::round(srcW * r);
    const int newH = (int)std::round(srcH * r);
    const int dw = (m_inW - newW) / 2;
    const int dh = (m_inH - newH) / 2;
    info.r = r;
    info.dw = dw;
    info.dh = dh;
    cv::Mat resized;

    cv::resize(srcBgr, resized, cv::Size(newW, newH), 0, 0, cv::INTER_LINEAR);
    cv::Mat out(m_inH, m_inW, CV_8UC3, cv::Scalar(114, 114, 114));
    resized.copyTo(out(cv::Rect(dw, dh, newW, newH)));

    return out;
}

void DetectorWorker::processFrame(const QImage &img)
{
    if (!m_ready || img.isNull())
        return;

    bool shouldStart = false;
    {
        QMutexLocker lock(&m_mutex);
        if (!m_busy) {
            m_busy = true;
            m_current = img; // implicit-share
            shouldStart = true;
        } else {
            // Overwrite pending with newest
            m_pending = img;
        }
    }

    if (shouldStart) {
        // Run in worker thread event loop; yields between frames
        QMetaObject::invokeMethod(this, &DetectorWorker::doWork, Qt::QueuedConnection);
    }
}

void DetectorWorker::doWork()
{
    QImage img;
    {
        QMutexLocker lock(&m_mutex);
        img = m_current;
    }

    if (!m_ready || img.isNull()) {
        QMutexLocker lock(&m_mutex);
        m_busy = false;

        return;
    }

    // Convert & preprocess
    const cv::Mat bgrSrc = qimageToBgrMat(img);

    LetterboxInfo lb;
    cv::Mat bgrInput;
    if (m_useLetterbox) {
        bgrInput = letterbox(bgrSrc, lb);
    } else {
        lb = LetterboxInfo{};
        cv::resize(bgrSrc, bgrInput, cv::Size(m_inW, m_inH));
    }

    // Blob: swapRB=true => BGR->RGB
    cv::Mat blob = cv::dnn::blobFromImage(
        bgrInput, 1.0/255.0, cv::Size(m_inW, m_inH),
        cv::Scalar(), true, false
        );

    m_net.setInput(blob);
    std::vector<cv::Mat> outs;

try {
        m_net.forward(outs, m_net.getUnconnectedOutLayersNames());
    } catch (...) {
        outs.clear();
        outs.push_back(m_net.forward());
    }

    QVector<Det> dets;
    if (!outs.empty())
        dets = runYoloFromOutput(outs[0], img.width(), img.height(), lb);

    // Emit text
    QStringList lines;
    lines.reserve(dets.size());
    for (const auto &d : dets) {
        const QString label =
            (d.classId >= 0 && d.classId < m_labels.size()) ? m_labels[d.classId] : QString("class_%1").arg(d.classId);

        lines << QString("%1 score=%2 box=[%3,%4,%5,%6]")
                     .arg(label)
                     .arg(d.score, 0, 'f', 2)
                     .arg(d.box.x(), 0, 'f', 0)
                     .arg(d.box.y(), 0, 'f', 0)
                     .arg(d.box.width(), 0, 'f', 0)
                     .arg(d.box.height(), 0, 'f', 0);
    }

    emit detectionsTextReady(lines);
    emit annotatedFrameReady(drawDetections(img, dets));

    // Schedule next if pending exists
    bool hasNext = false;
    {
        QMutexLocker lock(&m_mutex);
        if (!m_pending.isNull()) {
            m_current = m_pending;
            m_pending = QImage();
            hasNext = true;
        } else {
            m_busy = false;
        }
    }

    if (hasNext) {
        QMetaObject::invokeMethod(this, &DetectorWorker::doWork, Qt::QueuedConnection);
    }
}

QVector<DetectorWorker::Det> DetectorWorker::runYoloFromOutput(const cv::Mat &outRaw, int imgW, int imgH, const LetterboxInfo &lb)
{
    // Ensure float32 for safe reading
    cv::Mat out;
    if (outRaw.depth() != CV_32F)
        outRaw.convertTo(out, CV_32F);
    else
        out = outRaw;

    // Debug shape (prints a few times)
    static int dbg = 0;
    if (dbg < 5) {
        dbg++;
        QString shape = "dims=" + QString::number(out.dims);

        for (int i = 0; i < out.dims; ++i)
            shape += " " + QString::number(out.size[i]);
        qDebug() << "ONNX output:" << shape << "type=" << out.type();

        cv::Mat flat = out.reshape(1, 1);
        double mn=0, mx=0;
        cv::minMaxLoc(flat, &mn, &mx);
        qDebug() << "output min/max:" << mn << mx;
    }

    QVector<Det> candidates;
    auto clamp01orPixels = [&](float &x, float &y, float &w, float &h, int refW, int refH) {
        if (x <= 1.5f && y <= 1.5f && w <= 1.5f && h <= 1.5f) {
            x *= (float)refW;
            y *= (float)refH;
            w *= (float)refW;
            h *= (float)refH;
        }
    };

    auto mapXYXYFromInputToImage = [&](float &x1, float &y1, float &x2, float &y2) {
        if (m_useLetterbox) {
            x1 = (x1 - (float)lb.dw) / lb.r;
            y1 = (y1 - (float)lb.dh) / lb.r;
            x2 = (x2 - (float)lb.dw) / lb.r;
            y2 = (y2 - (float)lb.dh) / lb.r;
        } else {
            x1 = x1 * imgW / (float)m_inW;
            x2 = x2 * imgW / (float)m_inW;
            y1 = y1 * imgH / (float)m_inH;
            y2 = y2 * imgH / (float)m_inH;
        }

        x1 = std::max(0.0f, std::min(x1, (float)imgW - 1.0f));
        y1 = std::max(0.0f, std::min(y1, (float)imgH - 1.0f));
        x2 = std::max(0.0f, std::min(x2, (float)imgW - 1.0f));
        y2 = std::max(0.0f, std::min(y2, (float)imgH - 1.0f));
    };

    auto addCandidateXYWH = [&](float cx, float cy, float w, float h, float score, int cls) {
        if (score < m_confThres)
            return;

        clamp01orPixels(cx, cy, w, h, m_inW, m_inH);

        float x1 = cx - w * 0.5f;
        float y1 = cy - h * 0.5f;
        float x2 = cx + w * 0.5f;
        float y2 = cy + h * 0.5f;

        mapXYXYFromInputToImage(x1, y1, x2, y2);

        const float ww = std::max(0.0f, x2 - x1);
        const float hh = std::max(0.0f, y2 - y1);

        if (ww <= 1.0f || hh <= 1.0f)
            return;

        candidates.push_back({ QRectF(x1, y1, ww, hh), cls, score });
    };

    // Your model prints: dims=3 1 84 2100  -> treat as [1, C, N] with C=84 (4 + 80)
    if (out.dims == 3 && out.size[0] == 1 && out.size[1] >= 6 && out.size[2] > 100) {
        const int C = out.size[1];
        const int N = out.size[2];
        const float *p = (const float*)out.data;
        const int labelCount = m_labels.size();
        const int clsStart = 4;
        const int clsEnd = (labelCount > 0) ? std::min(C, clsStart + labelCount) : C;

        // Print best-score range once to help tuning threshold
        static bool printedMax = false;
        float globalMax = 0.f;
        int globalIdx = -1;

        for (int i = 0; i < N; ++i) {
            float cx = p[0*N + i];
            float cy = p[1*N + i];
            float w  = p[2*N + i];
            float h  = p[3*N + i];
            int bestCls = -1;
            float bestScore = 0.f;         

       for (int c = clsStart; c < clsEnd; ++c) {
                float s = scoreToProb(p[c*N + i]);
                if (s > bestScore) { bestScore = s; bestCls = c - clsStart; }
            }

            if (!printedMax && bestScore > globalMax) {
                globalMax = bestScore;
                globalIdx = bestCls;
            }

            addCandidateXYWH(cx, cy, w, h, bestScore, bestCls);
        }

        if (!printedMax) {
            printedMax = true;
            qDebug() << "Max class score seen (first frame):" << globalMax
                     << "classIndex=" << globalIdx
                     << "(try confThres ~0.15-0.30)";
        }
    }

    // Fallback: [1, N, 5+classes]
    else if (out.dims == 3 && out.size[0] == 1 && out.size[2] >= 6) {
        const int N = out.size[1];
        const int C = out.size[2];
        const float *p = (const float*)out.data;
        const int labelCount = m_labels.size();
        const int clsStart = 5;
        const int clsEnd = (labelCount > 0) ? std::min(C, clsStart + labelCount) : C;

        for (int i = 0; i < N; ++i) {
            float cx = p[i*C + 0], cy = p[i*C + 1], w = p[i*C + 2], h = p[i*C + 3];
            float obj = scoreToProb(p[i*C + 4]);
            int bestCls = -1;
            float bestClsScore = 0.f;

            for (int c = clsStart; c < clsEnd; ++c) {
                float s = scoreToProb(p[i*C + c]);
                if (s > bestClsScore) { bestClsScore = s; bestCls = c - clsStart; }
            }

            addCandidateXYWH(cx, cy, w, h, obj * bestClsScore, bestCls);
        }
    }
    else {
        qWarning() << "Unsupported output shape. dims=" << out.dims;
        return {};
    }

    // NMS
    std::vector<cv::Rect> boxes;
    std::vector<float> scores;
    boxes.reserve((size_t)candidates.size());
    scores.reserve((size_t)candidates.size());

    for (const auto &d : candidates) {
        boxes.emplace_back((int)d.box.x(), (int)d.box.y(),
                           (int)d.box.width(), (int)d.box.height());
        scores.push_back(d.score);
    }

    std::vector<int> keep;
    cv::dnn::NMSBoxes(boxes, scores, m_confThres, m_nmsThres, keep);
    QVector<Det> finalDets;
    finalDets.reserve((int)keep.size());

    for (int idx : keep)
        finalDets.push_back(candidates[idx]);

    return finalDets;
}

QImage DetectorWorker::drawDetections(const QImage &src, const QVector<Det> &dets)
{
    QImage outImg = src.convertToFormat(QImage::Format_RGBA8888);

    QPainter p(&outImg);
    p.setRenderHint(QPainter::Antialiasing, true);

    QPen pen(Qt::green);
    pen.setWidth(3);
    p.setPen(pen);

    QFont f = p.font();
    f.setPointSize(32);
    f.setBold(true);
    p.setFont(f);

    for (const auto &d : dets) {
        const QString label =
            (d.classId >= 0 && d.classId < m_labels.size()) ? m_labels[d.classId] : QString("class_%1").arg(d.classId);

        p.drawRect(d.box);
        const QString text = QString("%1 %2").arg(label).arg(d.score, 0, 'f', 2);
        QRectF tr(d.box.x(), d.box.y() - 42, 320, 42);
        p.fillRect(tr, QColor(0, 0, 0, 160));
        p.setPen(Qt::white);
        p.drawText(tr.adjusted(6, 0, 0, 0), text);
        p.setPen(pen);
    }

    return outImg;
}

#pragma once
#include <QQuickImageProvider>
#include <QImage>
#include <QMutex>

class AnnotatedImageProvider : public QQuickImageProvider
{
public:
    AnnotatedImageProvider();

    // Called from your detector/worker when a new annotated frame is ready
    void setImage(const QImage &img);

    QImage requestImage(const QString &id, QSize *size, const QSize &requestedSize) override;

private:
    QImage m_img;
    QMutex m_mutex;
};

#include "AnnotatedImageProvider.h"
#include <QPainter>

AnnotatedImageProvider::AnnotatedImageProvider()
    : QQuickImageProvider(QQuickImageProvider::Image)
{
}

void AnnotatedImageProvider::setImage(const QImage &img)
{
    QMutexLocker lock(&m_mutex);
    m_img = img;
}

QImage AnnotatedImageProvider::requestImage(const QString &id, QSize *size, const QSize &requestedSize)
{
    // IMPORTANT: id may include query string, e.g. "frame?ts=123"
    QString baseId = id;
    const int q = baseId.indexOf('?');

    if (q >= 0)
        baseId.truncate(q);

    QImage out;
    {
        QMutexLocker lock(&m_mutex);
        out = m_img;
    }

    // If no image yet, return a placeholder instead of failing
    if (out.isNull()) {
        const int w = requestedSize.width()  > 0 ? requestedSize.width()  : 640;
        const int h = requestedSize.height() > 0 ? requestedSize.height() : 480;
        out = QImage(w, h, QImage::Format_RGBA8888);
        out.fill(QColor(20, 20, 20));

        QPainter p(&out);
        p.setPen(Qt::white);
        p.drawText(out.rect(), Qt::AlignCenter, "Waiting for frames...");
    }

    if (size)
        *size = out.size();

    // If you ever want to support multiple ids, check baseId here:
    // if (baseId != "frame") { ... }

    return out;
}

#include <QGuiApplication>
#include <QQmlApplicationEngine>
#include <QQmlContext>
#include <QCoreApplication>
#include <QDebug>
#include <QFile>
#include <QDir>
#include <QStandardPaths>
#include <QThread>
#include <QPermission>
#include <QCamera>
#include <QCameraDevice>
#include <QMediaCaptureSession>
#include <QMediaDevices>
#include "FrameGrabber.h"
#include "DetectorWorker.h"
#include "AnnotatedImageProvider.h"

extern int qInitResources_app_assets();
extern int qCleanupResources_app_assets();

static QString extractResourceToFile(const QString &qrcPath, const QString &targetName)
{
    QString dir = QStandardPaths::writableLocation(QStandardPaths::CacheLocation);
    if (dir.isEmpty())
        dir = QStandardPaths::writableLocation(QStandardPaths::AppDataLocation);

    if (dir.isEmpty()) {
        qCritical() << "No writable location for extracted assets.";

        return {};
    }

    QDir().mkpath(dir);
    const QString outPath = QDir(dir).filePath(targetName);

    QFile in(qrcPath);
    if (!in.exists()) {
        qCritical() << "Resource does not exist:" << qrcPath;

        return {};
    }

    if (!in.open(QIODevice::ReadOnly)) {
        qCritical() << "Failed to open resource:" << qrcPath << in.errorString();

        return {};
    }

    QFile out(outPath);
    if (!out.open(QIODevice::WriteOnly | QIODevice::Truncate)) {
        qCritical() << "Failed to open output:" << outPath << out.errorString();

        return {};
    }

    const QByteArray data = in.readAll();
    out.write(data);
    out.close();
    in.close();

    qDebug() << "Extracted" << qrcPath << "->" << outPath << "bytes=" << data.size();

    return outPath;
}

int main(int argc, char *argv[])
{
    QGuiApplication app(argc, argv);
    qInitResources_app_assets();
    QQmlApplicationEngine engine;

    auto *provider = new AnnotatedImageProvider();
    engine.addImageProvider("annotated", provider);

    FrameGrabber grabber;
    engine.rootContext()->setContextProperty("frameGrabber", &grabber);

    // Worker thread
    QThread workerThread;
    auto *worker = new DetectorWorker();
    worker->moveToThread(&workerThread);

    QObject::connect(&app, &QCoreApplication::aboutToQuit, [&](){
        if (workerThread.isRunning()) {
            QMetaObject::invokeMethod(worker, &QObject::deleteLater, Qt::QueuedConnection);
            workerThread.quit();
            workerThread.wait();
        } else {
            delete worker;
        }

        qCleanupResources_app_assets();
    });

    // Pipeline
    QObject::connect(&grabber, &FrameGrabber::frameReady,
                     worker, &DetectorWorker::processFrame,
                     Qt::QueuedConnection);

    QObject::connect(worker, &DetectorWorker::annotatedFrameReady,
                     &app, [provider](const QImage &img) {
                         provider->setImage(img);
                     },
                     Qt::QueuedConnection);

    QObject::connect(worker, &DetectorWorker::annotatedFrameReady,
                     &grabber, &FrameGrabber::onAnnotatedFrameReady,
                     Qt::QueuedConnection);

    // Print detections to logcat
    QObject::connect(worker, &DetectorWorker::detectionsTextReady,
                     &app, [](const QStringList &lines){
                         if (lines.isEmpty())
                             return;

                         qDebug().noquote() << "Detections:\n - " + lines.join("\n - ");
                     },
                     Qt::QueuedConnection);

    // Camera
    QMediaCaptureSession session;
    const QCameraDevice camDev = QMediaDevices::defaultVideoInput();

    if (camDev.isNull())
        qWarning() << "No default video input device found.";

    QCamera camera(camDev);
    session.setCamera(&camera);
    session.setVideoSink(grabber.videoSink());

    const QUrl url(QStringLiteral("qrc:/qml/Main.qml"));

    QObject::connect(&engine, &QQmlApplicationEngine::objectCreated,
                     &app, [&](QObject *obj, const QUrl &objUrl){
                         if (!obj && objUrl == url) {
                             qCritical() << "QML root object creation failed. Exiting.";
                             QCoreApplication::exit(-1);

                             return;
                         }

                         if (!obj)
                             return;

                         const QString modelPath = extractResourceToFile(":/assets/model.onnx", "model.onnx");

                         if (modelPath.isEmpty()) {
                             qCritical() << "Model extraction failed. Exiting.";
                             QCoreApplication::exit(-1);

                             return;
                         }

                         if (!workerThread.isRunning())         
                             workerThread.start();

                            QMetaObject::invokeMethod(worker, [worker](){
                             worker->setInputSize(320, 320);
                             worker->setUseLetterbox(true);
                             worker->setThresholds(0.20f, 0.45f); // lower for 320
                             worker->loadLabels(":/assets/labels.txt");
                         }, Qt::QueuedConnection);

                         QMetaObject::invokeMethod(worker, [worker, modelPath](){
                             worker->loadModel(modelPath);
                         }, Qt::QueuedConnection);

                         auto startCamera = [&](){
                             qDebug() << "Starting camera...";
                             camera.start();
                         };

                         QCameraPermission camPerm;
                         const auto status = app.checkPermission(camPerm);

                         if (status == Qt::PermissionStatus::Granted) {
                             startCamera();
                         } else {
                             app.requestPermission(camPerm, [&](const QPermission &perm){
                                 if (perm.status() == Qt::PermissionStatus::Granted)
                                     startCamera();
                                 else
                                     qWarning() << "Camera permission denied.";
                             });
                         }
                     },
                     Qt::QueuedConnection);

    engine.load(url);
    return app.exec();
}

ApplicationWindow {
    id: mainWnd
    visible: true
    width: 960
    height: 540
    title: "Qt Android AI Object Detection"

    Rectangle {
        anchors.fill: parent
        color: "#111"

        ColumnLayout {
            id: rootLayout
            anchors.fill: parent
            anchors.margins: 12
            spacing: 10

            Text {
                text: "Camera + ONNX Object Detection"
                color: "white"
                font.pixelSize: 18
                Layout.alignment: Qt.AlignHCenter
            }

            Item {
                Layout.fillWidth: true
                Layout.fillHeight: true
                clip: true

                Image {
                    anchors.fill: parent
                    cache: false
                    asynchronous: true
                    fillMode: Image.PreserveAspectFit
                    source: "image://annotated/frame?ts=" + frameGrabber.frameCounter
                }
            }

            Text {
                text: "Detections are printed to logcat (Qt debug output)."
                color: "#bbb"
                font.pixelSize: 12
                Layout.alignment: Qt.AlignHCenter
            }
        }
    }
}