example_matching_with_cuda.cpp

// these contains the descriptors from the two images
// so you need to convert the AliceVision data structure for descriptors into GpuMat
cv::cuda::GpuMat inputGPUDescriptors1;
cv::cuda::GpuMat inputGPUDescriptors2;

// match the features
// this creates a brute force matcher (it matches each feature against all the others in the other image (more efficient in GPU)
cv::Ptr<cv::cuda::DescriptorMatcher> matcher= cv::cuda::DescriptorMatcher::createBFMatcher();

// this contains for each descriptor a vector of size knn (in this case 2) with the best and second best match for the descriptor
std::vector<std::vector<cv::DMatch>> matchesknn;

const float ratioThreshold{.6f};

const int knn{2};
// do the knn match
matcher->knnMatch(inputGPUDescriptors1, inputGPUDescriptors2, matchesknn, knn);

// apply the ratio test
std::vector<cv::DMatch> matches;
matches.reserve(matchesknn.size());
for(const auto& m : matchesknn)
{
    // test the ratio between the distances of the best and second best, if lower than a threshold it add the best match
    if(m[0].distance < ratioThreshold*m[1].distance)
        matches.push_back(m[0]);
}
matches.shrink_to_fit();