Question

Does anyone happen to know why the OpenCV 2 DescriptorMatcher::radiusMatch() and knnMatch() take a vector<vector<DMatch>>& matches? I’m a bit confused about why it wouldn’t just a vector, since it’s just a single array of points in the scene that correspond to the training image, right?

I’ve got something like this:

void getMatchingPoints(
    const vector<vector<cv::DMatch> >& matches,
    const vector<cv::KeyPoint>& keyPtsTemplates,
    const vector<cv::KeyPoint>& keyPtsScene,
    vector<Vec2f>& ptsTemplate,
    vector<Vec2f>& ptsScene
    )
{
    ptsTemplate.clear();
    ptsScene.clear();

    for (size_t k = 0; k < matches.size(); k++)
    {
        for (size_t i = 0; i < matches[k].size(); i++)
        {
            const cv::DMatch& match = matches[k][i];
            ptsScene.push_back(fromOcv(keyPtsScene[match.queryIdx].pt));
            ptsTemplate.push_back(fromOcv(keyPtsTemplates[match.trainIdx].pt));
        }
    }
}

but I’m a bit confused about how to actually map the approx. location of the object once I have them all in ptsScene. The points seem scattered to me when I just draw them, so I think I’m missing what the nested vectors represent.

Answer 1

The knnMatch function will return the k nearest-neighbour matches, i.e. if you call knnMatch(queryDescriptors, trainDescriptors, matchesQueryToTrain, 3) where in this case k=3, then for each training point, it will find the 3 best matches from the query set.

In terms of your vector<vector<DMatch>>, this means that the outer vector is a vector of each query->train match, and the inner vector is a vector of your k nearest matches.

There is quite a good example of how to use these k matches along with a cross-checking method in this other question.

If you want a simple 1-1 matching, then you can call knnMatch with k=1 which will return an inner vector of size 1, or just call match which outputs matches in the form vector<DMatch> with no second vector.