Add LOGOS features matching strategy

[catree]

Merge with extra: opencv/opencv_extra#690

This pullrequest changes

Original code: LOGOS

Regression tests consist in comparing the output of the current implementation w.r.t. the original Matlab code, see opencv/opencv_extra#690

Note:

• pointers are used as in the original code. I did not manage to use cv::Ptr successfully (got some memory leak issues)
• returned distance for the matches will be zero
• LOGOS parameters cannot be set
• parallelization could be used
• main changes should be code convention and float type instead of double

---

This method requires to build a BoW for a considered image dataset. Something like:

        std::string pattern = "Oxford City Centre/Images/*.jpg";
        vector<String> results;
        glob(pattern, results);

        Ptr<Feature2D> surf = SURF::create();
        vector<KeyPoint> bowKeypoints;
        Mat bowDescriptors;
        for (size_t i = 0; i < results.size(); i++)
        {
            Mat img = imread(results[i]);

            std::vector<KeyPoint> keypoints;
            Mat descriptors;
            surf->detectAndCompute(img, noArray(), keypoints, descriptors);

            bowKeypoints.insert(bowKeypoints.end(), keypoints.begin(), keypoints.end());
            bowDescriptors.push_back(descriptors);
        }

        int clusterCount = 512;
        BOWKMeansTrainer bowTrainer(clusterCount);
        Mat bowCenters = bowTrainer.cluster(bowDescriptors);

        FileStorage file("Oxford_BOW_512.yml", FileStorage::WRITE);
        file << "matName" << bowCenters;
    }

Cosine distance can be used to get the nearest BoW cluster for each descriptor.

[alalek]

Thank you for contribution!

[alalek]

Please avoid using of legacy "cvtest::BaseTest". It is not needed after migration on GoogleTest. Just write:

TEST(XFeatures2d_LogosMatcher, logos_matcher_regression)
{
    const std::string vP1_path ...
    ...
}

ts->get_data_path() + "..."

Please use cvtest::findDataFile(...) instead.

[alalek]

ASSERT_EQ(static_cast<int>(groundTruth.size()), correctMatches)
    << "groundTruth: " << groundTruth.size()
    << " ; matchesLogos: " << matchesLogos.size()
    << " ; correctMatches: " << correctMatches
;

[alalek]

Could you please backport sorting onto 3.4 branch too?
With some note which order should be preserved (like "cite id" string)

[alalek]

Is this file code covered by 3rdparty license too? Is where some adopted code or it has been written from scratch ?

[catree]

I have removed the license.

[alalek]

Well done! Thank you 👍