kalwalt
commented
2 years ago Speaking about sorting the keypoints, OpenCV use nor qsort or sort but anohter sorting algorithm, probably because more efficient ? KeyPoint OpenCV class and Keypoint_t from jsfeat are almost similar, and so probably we can adapt the code from OpenCV to our needs. If we look the Orb code in ComputeKeypoints method keypoints are removed those close to the border and then filtered:
// Remove keypoints very close to the border
KeyPointsFilter::runByImageBorder(keypoints, img.size(), edgeThreshold);
// Keep more points than necessary as FAST does not give amazing corners
KeyPointsFilter::retainBest(keypoints, scoreType == ORB_Impl::HARRIS_SCORE ? 2 * featuresNum : featuresNum);this is the retainBest method:
// takes keypoints and culls them by the response
void KeyPointsFilter::retainBest(std::vector<KeyPoint>& keypoints, int n_points)
{
//this is only necessary if the keypoints size is greater than the number of desired points.
if( n_points >= 0 && keypoints.size() > (size_t)n_points )
{
if (n_points==0)
{
keypoints.clear();
return;
}
//first use nth element to partition the keypoints into the best and worst.
std::nth_element(keypoints.begin(), keypoints.begin() + n_points - 1, keypoints.end(), KeypointResponseGreater());
//this is the boundary response, and in the case of FAST may be ambiguous
float ambiguous_response = keypoints[n_points - 1].response;
//use std::partition to grab all of the keypoints with the boundary response.
std::vector<KeyPoint>::const_iterator new_end =
std::partition(keypoints.begin() + n_points, keypoints.end(),
KeypointResponseGreaterThanOrEqualToThreshold(ambiguous_response));
//resize the keypoints, given this new end point. nth_element and partition reordered the points inplace
keypoints.resize(new_end - keypoints.begin());
}
}
https://github.com/opencv/opencv/blob/64aad34cb4abfb6a2603f3f4ecae7f4f0ba1414d/modules/features2d/src/keypoint.cpp#L68-L90 May we implement this inside webarkit-jsfeat-cpp? we have to try!
In this Pull request i will try do develop some utilities to train a Orb pattern. Partially following the jsfeat code but also OpenCV implementation as it was the original one.