OpenCV Why does number of SIFT keypoints go over threshold?

edit

I think that your are right @LBerger. I tested with this image and this code:

  cv::Ptr<cv::Feature2D> sift = cv::xfeatures2d::SIFT::create(100);    
  sift->detectAndCompute(matImg, cv::Mat(), kpts, descriptors);
  if(kpts.size() > 100) {
    std::cout << "kpts=" << kpts.size() << std::endl;
    std::map<float, std::vector<cv::KeyPoint> > map;
    for(size_t i = 0; i < kpts.size(); i++) {
      map[kpts[i].response].push_back(kpts[i]);
      std::cout << "response=" << std::fixed << std::setprecision(9) << kpts[i].response << std::endl;
    }

    std::cout << "map=" << map.size() << std::endl;
  }

I got 61 keypoints with different response and 101 in total.

That sounds like a bug to me. In this case algorithm should just ignore other features. Accorfing documentation it supposes to keep numer of best features.

I don't think it's a bug but if you think so you can make an issue or improve doc using a pull request. You are welcome

It is not a bug, but as suggested due to the fact that keypoints can have the same score or strength. It can therefore only limit it to the closest correct amount. What you should do is physically limit the output afterwards. It will never be 99, always 100 or more.

That's what I did, thanks.