Add functionality to save inference images for the debugging purposes.

Save bmp images of inference results to /tmp as bmp files. BMP was chosen to reduce encoding time. Saving is fully threaded. It can be enable with qmake CONFIG+=save_images option Also: - use antialised fonts in RKNN inference - moved class strings to inference base class - fixed silly segfault in ONNX inference - prevent writing results if class if exceeds valid values Issue: https://denyspopov.atlassian.net/browse/AZ-38 Type: Improvement
2026-04-22 09:16:35 +00:00 · 2024-08-19 12:15:42 +03:00
parent 022e4a1200
commit be59a02f9b
9 changed files with 182 additions and 104 deletions
@@ -2,6 +2,7 @@
 #include <opencv2/highgui.hpp>
 #include "aiengine.h"
 #include "aiengineinference.h"
+#include "aiengineimagesaver.h"

 #if defined(OPI5_BUILD)
 #include "src-opi5/aiengineinferenceopi5.h"
@@ -82,6 +83,13 @@ void AiEngine::inferenceResultsReceivedSlot(AiEngineInferenceResult result)
    }

    cv::imshow("Received Frame", result.frame);
+
+#ifdef SAVE_IMAGES
+    static int imageCounter = 0;
+    AiEngineImageSaver *saver = new AiEngineImageSaver(result.frame, ++imageCounter);
+    saver->start();
+    connect(saver, &AiEngineImageSaver::finished, saver, &QObject::deleteLater);
+#endif
 }


@@ -0,0 +1,31 @@
+#include <QFile>
+#include <QDataStream>
+#include <QDebug>
+#include "aiengineimagesaver.h"
+
+AiEngineImageSaver::AiEngineImageSaver(const cv::Mat &image, int imageNumber, QObject *parent)
+    : QThread(parent), image(image.clone()), imageNumber(imageNumber)
+{
+}
+
+AiEngineImageSaver::~AiEngineImageSaver()
+{
+    wait();
+}
+
+void AiEngineImageSaver::run()
+{
+    if (image.empty() || image.channels() == 0) {
+        qWarning() << "AiEngineImageSaver. Empty image or no channels, nothing to save.";
+        return;
+    }
+
+    // Calculate the size of the upper half
+    int halfHeight = image.rows / 2;
+    cv::Mat upperHalf = image(cv::Rect(0, 0, image.cols, halfHeight));
+
+    // Use bpm to reduce encoding time.
+    QString filePath = QString("/tmp/image-%1.bmp").arg(imageNumber, 5, 10, QChar('0'));
+
+    cv::imwrite(filePath.toStdString(), upperHalf);
+}
@@ -0,0 +1,21 @@
+#pragma once
+
+#include <QThread>
+#include <QString>
+#include <opencv2/opencv.hpp>
+
+class AiEngineImageSaver : public QThread
+{
+    Q_OBJECT
+
+public:
+    AiEngineImageSaver(const cv::Mat &image, int imageNumber, QObject *parent = nullptr);
+    ~AiEngineImageSaver();
+
+protected:
+    void run() override;
+
+private:
+    cv::Mat image;
+    int imageNumber;
+};
@@ -5,7 +5,105 @@ AiEngineInference::AiEngineInference(QString modelPath, QObject *parent)
    : QObject{parent},
      mModelPath(modelPath),
      mActive(false)
-{}
+{
+#ifdef YOLO_ONNX
+    mClassNames = {
+        "person",
+        "bicycle",
+        "car",
+        "motorcycle",
+        "airplane",
+        "bus",
+        "train",
+        "truck",
+        "boat",
+        "traffic light",
+        "fire hydrant",
+        "stop sign",
+        "parking meter",
+        "bench",
+        "bird",
+        "cat",
+        "dog",
+        "horse",
+        "sheep",
+        "cow",
+        "elephant",
+        "bear",
+        "zebra",
+        "giraffe",
+        "backpack",
+        "umbrella",
+        "handbag",
+        "tie",
+        "suitcase",
+        "frisbee",
+        "skis",
+        "snowboard",
+        "sports ball",
+        "kite",
+        "baseball bat",
+        "baseball glove",
+        "skateboard",
+        "surfboard",
+        "tennis racket",
+        "bottle",
+        "wine glass",
+        "cup",
+        "fork",
+        "knife",
+        "spoon",
+        "bowl",
+        "banana",
+        "apple",
+        "sandwich",
+        "orange",
+        "broccoli",
+        "carrot",
+        "hot dog",
+        "pizza",
+        "donut",
+        "cake",
+        "chair",
+        "couch",
+        "potted plant",
+        "bed",
+        "dining table",
+        "toilet",
+        "tv",
+        "laptop",
+        "mouse",
+        "remote",
+        "keyboard",
+        "cell phone",
+        "microwave",
+        "oven",
+        "toaster",
+        "sink",
+        "refrigerator",
+        "book",
+        "clock",
+        "vase",
+        "scissors",
+        "teddy bear",
+        "hair drier",
+        "toothbrush"
+    };
+#else
+    mClassNames = {
+        "Armoured vehicle",
+        "Truck",
+        "Vehicle",
+        "Artillery",
+        "Shadow artillery",
+        "Trenches",
+        "Military man",
+        "Tyre tracks",
+        "Additional protection tank",
+        "Smoke"
+    };
+#endif
+}


 bool AiEngineInference::isActive(void)
@@ -40,6 +40,7 @@ protected:
    QString mModelPath;
    bool mActive;
    int mNumber;
+    QVector<QString> mClassNames;

 public slots:
    virtual void performInferenceSlot(cv::Mat frame) = 0;
@@ -16,9 +16,15 @@ else {
 }

 yolo_onnx {
+    message("Using YOLOv8 ONNX models for indoor testing.")
    QMAKE_CXXFLAGS += -DYOLO_ONNX
 }

+save_images {
+    message("Saving inference images to /tmp.")
+    QMAKE_CXXFLAGS += -DSAVE_IMAGES
+}
+
 opi5 {
    message("OPI5 build")
    CONFIG += link_pkgconfig
@@ -14,104 +14,7 @@ AiEngineInferencevOnnxRuntime::AiEngineInferencevOnnxRuntime(QString modelPath,
    mPredictor(modelPath.toStdString(), confThreshold, iouThreshold, maskThreshold)
 {
    qDebug() << "TUOMAS AiEngineInferencevOnnxRuntime() mModelPath=" << mModelPath;
-
-#ifdef YOLO_ONNX
-    mClassNames = {
-        "person",
-        "bicycle",
-        "car",
-        "motorcycle",
-        "airplane",
-        "bus",
-        "train",
-        "truck",
-        "boat",
-        "traffic light",
-        "fire hydrant",
-        "stop sign",
-        "parking meter",
-        "bench",
-        "bird",
-        "cat",
-        "dog",
-        "horse",
-        "sheep",
-        "cow",
-        "elephant",
-        "bear",
-        "zebra",
-        "giraffe",
-        "backpack",
-        "umbrella",
-        "handbag",
-        "tie",
-        "suitcase",
-        "frisbee",
-        "skis",
-        "snowboard",
-        "sports ball",
-        "kite",
-        "baseball bat",
-        "baseball glove",
-        "skateboard",
-        "surfboard",
-        "tennis racket",
-        "bottle",
-        "wine glass",
-        "cup",
-        "fork",
-        "knife",
-        "spoon",
-        "bowl",
-        "banana",
-        "apple",
-        "sandwich",
-        "orange",
-        "broccoli",
-        "carrot",
-        "hot dog",
-        "pizza",
-        "donut",
-        "cake",
-        "chair",
-        "couch",
-        "potted plant",
-        "bed",
-        "dining table",
-        "toilet",
-        "tv",
-        "laptop",
-        "mouse",
-        "remote",
-        "keyboard",
-        "cell phone",
-        "microwave",
-        "oven",
-        "toaster",
-        "sink",
-        "refrigerator",
-        "book",
-        "clock",
-        "vase",
-        "scissors",
-        "teddy bear",
-        "hair drier",
-        "toothbrush"
-    };
-#else
-    mClassNames = {
-        "Armoured vehicle",
-        "Truck",
-        "Vehicle",
-        "Artillery",
-        "Shadow artillery",
-        "Trenches",
-        "Military man",
-        "Tyre tracks",
-        "Additional protection tank",
-        "Smoke"
-    };
-#endif
+    qDebug() << "AiEngineInferencevOnnxRuntime() mClassNames.size() =" << mClassNames.size();
 }


@@ -183,6 +86,11 @@ void AiEngineInferencevOnnxRuntime::performInferenceSlot(cv::Mat frame)
        for (uint i = 0; i < detections.size(); i++) {
            const Yolov8Result &detection = detections[i];

+            if (detection.classId >= mClassNames.size()) {
+                qDebug() << "performInferenceSlot() invalid classId =" << detection.classId;
+                continue;
+            }
+
            // Add detected objects to the results
            AiEngineObject object;
            object.classId          = detection.classId;
@@ -17,5 +17,4 @@ public slots:
 private:
    cv::Mat drawLabels(const cv::Mat &image, const std::vector<Yolov8Result> &detections);
    YOLOPredictor mPredictor;
-    QVector<QString> mClassNames;
 };
@@ -90,6 +90,12 @@ void AiEngineInferenceOpi5::drawObjects(cv::Mat& image, const object_detect_resu
    for (int i = 0; i < result_list.count; i++) {
        const object_detect_result& result = result_list.results[i];

+        if (result.cls_id >= mClassNames.size()) {
+            continue;
+        }
+
+        fprintf(stderr, "TUOMAS [%d] prop = %f\n", i, result.prop);
+
        int left = result.box.left;
        int top = result.box.top;
        int right = result.box.right;
@@ -99,9 +105,10 @@ void AiEngineInferenceOpi5::drawObjects(cv::Mat& image, const object_detect_resu

        // Text
        char c_text[256];
-        sprintf(c_text, "%s %.1f%%", coco_cls_to_name(result.cls_id), result.prop * 100);
+        //sprintf(c_text, "%s %d%%", coco_cls_to_name(result.cls_id), (int)(round(result.prop * 100)));
+        sprintf(c_text, "%s %d%%", mClassNames[result.cls_id].toStdString().c_str(), (int)(round(result.prop * 100)));
        cv::Point textOrg(left, top - 5);
-        cv::putText(image, std::string(c_text), textOrg, cv::FONT_HERSHEY_SIMPLEX, result.prop, cv::Scalar(0, 0, 255), 1);
+        cv::putText(image, std::string(c_text), textOrg, cv::FONT_HERSHEY_COMPLEX, result.prop, cv::Scalar(0, 0, 255), 1, cv::LINE_AA);
    }
 }

@@ -117,6 +124,7 @@ void AiEngineInferenceOpi5::performInferenceSlot(cv::Mat frame)

    object_detect_result_list od_results;
    int ret = inference_yolov8_model(&mRrknnAppCtx0, &imgBuffer, &od_results, mNumber);
+    freeImageBuffer(imgBuffer);
    if (ret != 0) {
        qDebug() << "AiEngineInferenceOpi5::performInferenceSlot() failure! ret: " << ret;
        mActive = false;
@@ -138,8 +146,6 @@ void AiEngineInferenceOpi5::performInferenceSlot(cv::Mat frame)
    }

    drawObjects(scaledFrame, od_results);
-    freeImageBuffer(imgBuffer);
-
    result.frame = scaledFrame.clone();
    emit resultsReady(result);