Bob Woodley's profile - activity

I'm getting poor results using a variety of Feature Detection Algos on Android. I cannot use SURF, SIFT, or BRISK because they are too slow (I'm using a Samsung Galaxy S5). So I've worked with ORB Detector/Orb Extractor, as well as with Orb Detector/FREAK extractor, both of which perform in real-time. However they aren't very good at finding my original training object in any scene with even small amounts of clutter.

I filter by Hamming distance. I apply cross-check and ratio test. I apply KNN. I use various combinations of these. I can't get a clear signal out of the noise. This photo below probably indicates the highest quality feature matching I can get, and there are many points that are not correct. The high number of errors would prevent an algo from confidentally identifying the object.

Is it possible to get a higher quality result?

Possibly my code is wrong, so here are the key lines of code, extracted from a larger class. In this example I'm using ORB Feature Detector and Descriptor Extractor with BF HAMMINGLUT matcher, and a simple distance filter.

DescriptorExtractor descriptorExtractor;
DescriptorMatcher _matcher;
FeatureDetector _detector;

Mat _descriptors;
MatOfKeyPoint _keypoints;

Mat _descriptors2;
MatOfKeyPoint _keypoints2;

private Mat trainFeatureDetector(CvCameraViewFrame inputFrame) {
    Mat gray1 = inputFrame.gray();

    _descriptors = new Mat();
    _keypoints = new MatOfKeyPoint();

    _detector = FeatureDetector.create(FeatureDetector.ORB);
    _matcher = DescriptorMatcher.create(DescriptorMatcher.BRUTEFORCE_HAMMINGLUT);
    _detector.detect(gray1, _keypoints, _descriptors);

    descriptorExtractor = DescriptorExtractor.create(DescriptorExtractor.ORB);
    descriptorExtractor.compute(gray1, _keypoints, _descriptors);
}
public Mat onCameraFrame(CvCameraViewFrame inputFrame) {
    Mat gray2 = inputFrame.gray();
    _descriptors2 = new Mat();
    _keypoints2 = new MatOfKeyPoint();

    _detector.detect(gray2, _keypoints2);
    descriptorExtractor.compute(gray2, _keypoints2, _descriptors2);

    MatOfDMatch matches12 = new MatOfDMatch();
    _matcher.match(_descriptors, _descriptors2, matches12);
    List<DMatch> matches12_list = matches12.toList();

    // .... filter based on distance... 
    //-- Quick calculation of min distances between keypoints
    double min_dist = 0;
    for (int i = 0; i < matches12_list.size(); i++) {
        double dist = matches12_list.get(i).distance;
        if (dist < min_dist) min_dist = dist;
    }

    //-- Draw only "good" matches (i.e. whose distance is less than 3*min_dist )
    List<DMatch> good_matches_list = new ArrayList<>();
    for (int i = 0; i < matches12_list.size(); i++) {
        if (matches12_list.get(i).distance < 3 * min_dist) {
            good_matches_list.add(matches12_list.get(i));
        }
    }

    drawMatches(gray2, _keypoints2, good_matches_list, (double) gray2.height(), (double) gray2.width());
}

Thanks for the explanation and the solution.

I have an app that needs to run in portrait mode and I want to have a small camera preview displayed on the screen. Opencv defaults to landscape for preview. The following code rotates it back to portrait:

 public Mat onCameraFrame(CameraBridgeViewBase.CvCameraViewFrame inputFrame) {
    mRgba = inputFrame.rgba();

    Mat tempMat = mRgba.t();
    Core.flip(tempMat, mRgbaT, 1);
    Imgproc.resize(mRgbaT, dst, mRgba.size());
    return dst;
}

The problem is that this code leaks memory and crashes after 10 or 15 seconds. I have tried various permutations of using release and using temporary matrices. It always seems to crash on the resize() function so I suppose that is where the leak is.

For the above example these are the declarations:

private Mat mRgba;
private Mat mRgbaT;
private Mat dst;

initialized in onCameraViewStarted():

public void onCameraViewStarted(int width, int height) {
    mRgbaT = new Mat();
    dst = new Mat();
}

I need help managing the memory correctly. I'd also like to understand what the problem is. Since Java is garbage collected it feels mighty strange to be required to call release().

I have built a debug x64 VS2010 project containing the facerec_demo.cpp class from the samples directory. I am use the ATT database with the supplied csv file from the samples directory.

I am using OpenCV2.4.2. Downloaded 4 days ago.

I get: Unhandled exception at 0x000007fef18e26bb in FaceRecon.exe: 0xC0000005: Access violation reading location 0xffffffffffffffff

at this line:

int predictedLabel = model->predict(testSample);

Any ideas? Has this been tested on 64 bit debug mode?

Thanks!