Eliminating false positives in feature descriptors

Question

Given two occupancy grid maps converted into cv::Mat grayscale, I used ORB feature descriptor and match the extracted features from both images. Most of it are false positives. Here below, I only showed one relation to show that it is really a false positive.

enter image description here

Ultimately, what I want to do is to find the right transformation between the two occupancy grid maps in order to merge their maps into one globally consistent map. The current code that I have is shown below:

  // Load the two maps
    nav_msgs::OccupancyGrid map1;
    initOccupancyGridMap1(map1);
    nav_msgs::OccupancyGrid map2;
    initOccupancyGridMap2(map2);

    // Convert the two maps to grayscale images
    cv::Mat im1 = cvtMapToMat(map1);
    cv::Mat im2 = cvtMapToMat(map2);

    // Feature Descriptor Extraction
    cv::OrbFeatureDetector featureDetector;
    cv::OrbDescriptorExtractor featureExtractor;
    std::vector<cv::KeyPoint> kp1;
    std::vector<cv::KeyPoint> kp2;
    cv::Mat d1;
    cv::Mat d2;
    std::vector<cv::DMatch> matches;
    cv::BFMatcher dematc(cv::NORM_HAMMING, false);
    // 1. Detect keypoints
    featureDetector.detect(im1, kp1);
    featureDetector.detect(im2, kp2);
    // 2. Extract descriptors
    featureExtractor.compute(im1, kp1, d1);
    featureExtractor.compute(im2, kp2, d2);
    // 3. Match keypoints
    dematc.match(d1, d2, matches);

    for (int i = 0; i < matches.size(); i++){
        std::vector<cv::DMatch> match(1,matches[i]);

        std::cout << "Distance: " << match[0].distance << std::endl;
        cv::Mat img_show;
        drawMatches(im1, kp1, im2, kp2, match, img_show);
        imshow("Matches", img_show);
        cv::waitKey(0);
    }

possible duplicate of [improve matching of feature points with openCV](http://stackoverflow.com/questions/17967950/improve-matching-of-feature-points-with-opencv) — Aurelius, May 02 '14 at 20:19

score 0 · Answer 1 · edited Oct 29 '14 at 15:02

In order to remove the points that are very similar, I used the second nearest neighbor. If the point match with one point that is almost equal to other, you can't be sure that the match is correct.

For example, if your point A match with points B and C, you can consider B a good match if they differ at least 25% (that diference is what I chose, but you can try with other numbers). .

The code looks like:

 // 2. Extract descriptors
featureExtractor.compute(im1, kp1, d1);
featureExtractor.compute(im2, kp2, d2);
// 3. Match keypoints

std::vector<cv::DMatch> goodMatches;
// Instead match only one point, try to match with 2.
matcher.knnMatch( d1, d2, matches, 2 );

for(int i = 0; i< matches.size(); i++){  // matches from matcher
    //with this condition we check that the points are at least 25% diferent.
    if(matches[i].at(0).distance< 0.75*matches[i].at(1).distance){
         goodMatches.push_back(matches[i].at(0));
    }
}

Eliminating false positives in feature descriptors

1 Answers1