Ask Your Question
0

Inaccurate feature matching

asked 2014-03-28 15:55:57 -0600

glethien gravatar image

updated 2014-03-28 20:14:32 -0600

Moster gravatar image

Hi, I am currently developing an AndroidApp using OpenCV4Android. The aim of this app is a structure form motion in a little larger scale. The concept of the app and the mathematical background is done. The problem is the matching of the feature points. But for a better understanding I will provide a little bit more background.

The problem is that I am having two images of the same scene. Between the images the camera has been translated and maybe there is some rotation. For the structure form motion I need the fundamental matrix and get the essential matrix for later usage.

As OpenCV4Android does not provide SIFT or SURF Feature matching, I am using ORB and Brutforce_Hamming matcher. The Featuredetector is getting excactly 500 Featurepoints per image (8MP) But the matcher for the featuerpoints is only gettin between 0 and 3 matches... I need far more matches for getting a good fundamental matrix.

I am doing the whole process on undistored images using the cameracalibration done before.

Here is the code I am using to get the matches and the FeaturePoints: // Create a feature detector which uses SIFT Features FeatureDetector detector = FeatureDetector.create(FeatureDetector.ORB);

// Create an extractor for the description of the feature points using 3 // WHAT IS 3
DescriptorExtractor extractor = DescriptorExtractor.create(DescriptorExtractor.ORB);

// Matches the described feature points of both images together using bruteforce_hamming
DescriptorMatcher matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_HAMMING);

// Minimal distance between two matches
double min_dist = 80; 
// Maximal distance between two matches
double max_dist = 10000;
/**
 * Calculates the matches of KeyPoints between two images
 * @Param left the left image
 * @Param right the right image
 * @return List<DMatch> List of matches between left and right
 */
private List<DMatch> computeGoodMatches(Mat left, Mat right, MatOfPoint2f object_left, MatOfPoint2f object_right){
    // For the left image

    MatOfKeyPoint key_left = new MatOfKeyPoint();
    detector.detect(left, key_left);

    publishProgress("Keypoints left: " + String.valueOf(key_left.size()));

    Mat desc_left = new Mat();
    extractor.compute(left, key_left, desc_left);

    // For the right image

    MatOfKeyPoint key_right = new MatOfKeyPoint();
    detector.detect(right, key_right);

    publishProgress("Keypoints right: " + String.valueOf(key_right.size()));

    Mat desc_right = new Mat();
    extractor.compute(right, key_right, desc_right);


    // Calculate the matches (evaluating the good matches happens later in the code)
    MatOfDMatch matches = new MatOfDMatch();
    matcher.match(desc_left, desc_right, matches);

    List<DMatch> matchesList = matches.toList();

    publishProgress("Number matches: " + String.valueOf(matchesList.size()));

    LinkedList<DMatch> good_matches = new LinkedList<DMatch>();

    // Quick calculation of min and max distance between matches 
    for( int j = 0; j < desc_left.rows(); j++ ){
        Double dist = (double) matchesList.get(j).distance;
        if( dist < min_dist ) min_dist = dist;
        if( dist > max_dist ) max_dist = dist;
    }

    // Only use good matches
    // Good = 2*min_dist or 0.02 
    for(int j = 0; j < desc_left.rows(); j++){
        if(matchesList.get(j).distance <= min_dist){
            good_matches.addLast(matchesList.get(j));
        }
    }


    List<KeyPoint> list_key_left =  key_left.toList();
    List<KeyPoint> list_key_right = key_right.toList();

    LinkedList<Point> objList1 = new LinkedList<Point>();
    LinkedList<Point> objList2 = new LinkedList<Point>();

    for(int j = 0; j<good_matches.size(); j++){
        objList1.addLast(list_key_left.get(good_matches.get(j).queryIdx).pt);
        objList2.addLast(list_key_right.get(good_matches.get(j).trainIdx).pt);
    }

    object_left.fromList(objList1);
    object_right ...
(more)
edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
2

answered 2014-03-28 20:10:53 -0600

Moster gravatar image

updated 2014-03-28 20:11:34 -0600

Your threshold for the "good features extraction" is only min_dist, but should be something like 2 or 3 times min_dist. Its kind of obvious that you only get a few matches if the distance should be equal to min_dist.

for(int j = 0; j < desc_left.rows(); j++){
    if(matchesList.get(j).distance <= 2*min_dist){
        good_matches.addLast(matchesList.get(j));
    }
}
edit flag offensive delete link more

Question Tools

Stats

Asked: 2014-03-28 15:55:57 -0600

Seen: 1,066 times

Last updated: Mar 28 '14