Open CV - Matching video frame with image

Question

I am trying to compare the key points of a video frame with an image I upload. This is my code. The img parameter is the uploaded picture, pixels_color is the pixels from the frame, height and width is also from the frame.

int Image::match(ofImage img){

    cv::SurfFeatureDetector detector(400);
    vector<cv::KeyPoint> keypoints1, keypoints2;
    cv::Mat img1(height, width, CV_8UC3, pixels_color);
    cv::Mat img2(img.getHeight(), img.getWidth(), CV_8UC3, img.getPixels());

    detector.detect(img1, keypoints1);
    detector.detect(img2, keypoints2);

    cv::SurfDescriptorExtractor extractor;
    cv::Mat descriptors1, descriptors2;
    extractor.compute(img1, keypoints1, descriptors1);
    extractor.compute(img2, keypoints2, descriptors2);

    cv::BruteForceMatcher<cv::L2<float>> matcher;
    vector<cv::DMatch> matches;
    matcher.match(descriptors1, descriptors2, matches);

    return matches.size();
}

The problem is the return is always size of keypoints1 . I do not know why, unless my keypoints2 has size 0, the return will always be the size of keypoints1.

It does not matter if the ofImage img has something to do with the frame or not, it will always return the size of keypoints1. Here is an example:

Answer 1

I still didn't understand what happened, so I used a different approach, I used FlannBasedMatcher with a normalization, with permitted me to achieve what I wanted to:

int Image::match(ofImage img){

    bool result = false;
    float neares_neighbor_distance_ratio = 0.7f;

    cv::SurfFeatureDetector detector(400);
    vector<cv::KeyPoint> keypoints1, keypoints2;
    cv::SurfDescriptorExtractor extractor;
    cv::Mat descriptors1, descriptors2;

    cv::Mat img1(height, width, CV_8UC3, pixels_color);
    cv::Mat img2(img.getHeight(), img.getWidth(), CV_8UC3, img.getPixels());

    detector.detect(img1, keypoints1);
    detector.detect(img2, keypoints2);

    extractor.compute(img1, keypoints1, descriptors1);
    extractor.compute(img2, keypoints2, descriptors2);

    cv::FlannBasedMatcher matcher;
    vector<vector<cv::DMatch>> matches; 
    matcher.knnMatch(descriptors1, descriptors2, matches, 2);

    vector<cv::DMatch> good_matches;
    good_matches.reserve(matches.size());  

    for(size_t i = 0; i < matches.size(); ++i)
    { 
        if(matches[i].size() < 2)
            continue;

        const cv::DMatch &m1 = matches[i][0];
        const cv::DMatch &m2 = matches[i][1];

        if(m1.distance <= neares_neighbor_distance_ratio * m2.distance)        
            good_matches.push_back(m1);     
    }

    return good_matches.size();
}

the method is a bit bigger, but it worked I guess.