accurate confidence

[fatalfeel]

for (auto& detection : detections)
{
for (int i = 0; i < detection.rows; i++)
{
cv::Mat onerow = detection.row(i);
cv::Mat scores = onerow.colRange(5, onerow.cols);
cv::minMaxLoc(scores, 0, &maxscore, 0, &classIdPoint);
finalConf = onerow.at(4) * maxscore;

    //if (maxscore > threshold)
    if (finalConf > threshold)
    {
        int centerX = static_cast<int>(onerow.at<float>(0) * resize_3840_bgr.cols);
        int centerY = static_cast<int>(onerow.at<float>(1) * resize_3840_bgr.rows);
        int width   = static_cast<int>(onerow.at<float>(2) * resize_3840_bgr.cols);
        int height  = static_cast<int>(onero{
for (int i = 0; i < detection.rows; i++)
{
    cv::Mat onerow = detection.row(i);
    cv::Mat scores = onerow.colRange(5, onerow.cols);
    cv::minMaxLoc(scores, 0, &maxscore, 0, &classIdPoint);
    finalConf = onerow.at<float>(4) * maxscore;

    //if (maxscore > threshold)
    if (finalConf > threshold)
    {
        int centerX = static_cast<int>(onerow.at<float>(0) * resize_3840_bgr.cols);
        int centerY = static_cast<int>(onerow.at<float>(1) * resize_3840_bgr.rows);
        int width   = static_cast<int>(onerow.at<float>(2) * resize_3840_bgr.cols);
        int height  = static_cast<int>(onerow.at<float>(3) * resize_3840_bgr.rows);
        int left    = centerX - width   / 2;
        int top     = centerY - height  / 2;

        boxes.push_back(cv::Rect(left, top, width, height));
        //classIds.push_back(classId);
        //confidences.push_back((float)maxscore);
        confidences.push_back((float)finalConf);
    }
}

}w.at(3) * resize_3840_bgr.rows);
int left = centerX - width / 2;
int top = centerY - height / 2;

        boxes.push_back(cv::Rect(left, top, width, height));
        //classIds.push_back(classId);
        //confidences.push_back((float)maxscore);
        confidences.push_back((float)finalConf);
    }
}

}

In YOLO, the decision-making process involves:

Locating objects: Based on the objectness score (onerow.at<float>(4)).
Classifying objects: Based on the class confidence scores (onerow.at<float>(5) through onerow.at<float>(10)).

By multiplying the objectness score and the class confidence score, you get a combined measure (finalConf) that represents the likelihood of a specific class being present in the bounding box.
Example:

For a row with the following values:

[0.5, 0.5, 0.2, 0.4, 0.9, 0.1, 0.7, 0.2, 0.05, 0.03, 0.02]

Center X = 0.5 (50% of image width).
Center Y = 0.5 (50% of image height).
Width = 0.2 (20% of image width).
Height = 0.4 (40% of image height).
Objectness Score = 0.9 (90% confidence that there is an object in this box).
Class Scores:
    Class 0: 0.1
    Class 1: 0.7
    Class 2: 0.2
    Class 3: 0.05
    Class 4: 0.03
    Class 5: 0.02

If the highest class confidence score is 0.7 (for Class 1):

finalConf = 0.9 * 0.7 = 0.63.

This final confidence is compared to the threshold to decide whether to keep this detection.

//python
maxcore = scores[class_id]
finalConf = detection[4] * maxcore
#if maxcore > CONF_THRESH:
if finalConf > CONF_THRESH:
center_x = int(detection[0] * im_w)
center_y = int(detection[1] * im_h)
width = int(detection[2] * im_w)
height = int(detection[3] * im_h)
left = int(center_x - width / 2)
top = int(center_y - height / 2)
class_ids.append(int(class_id))
#confidences.append(float(maxcore))
confidences.append(float(finalConf))
boxes.append([left, top, width, height])