Ask Your Question

Is it possible to generate key points from points tracked optical flow?

asked 2013-03-19 07:37:53 -0600

axjensen gravatar image


I'm wondering if it is possible to generate KeyPoints from the 2D location tracked by a optical flow algorithm, such as calcOpticalFlowPyrLK?

I would like to eventually extract features at the 2D location of the tracked points, by they are merely 2D points and not KeyPoints, and therefor not suitable for OpenCV's DescriptorExtractor.

I could, of course, just initialize the KeyPoints from the 2D locations, but then I cannot initialize them with values. such as angle, size etc., which would influence the quality of the features.


edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2013-03-19 08:18:44 -0600

Guanta gravatar image

Several ideas:

  1. First I would try your idea and initialize the 2D locations and initially assume no rotation and a scale of 1.0. If you like you can then add more keypoints at the same locations but with different scale (see e.g. how it is done in the dense-detector: detectors.cpp, l.162) and / or rotation parameters and see how good the features will match / be useful for classification.

  2. The most feature-detectors first detect 'robust' keypoints like corners. And from these ones they compute the major orientation / scale. So you could look in the source code of your favorite descriptor and modify it such that it uses your points instead of the original points of interest and then continue processing.

  3. Maybe this is possible for you: Take a window around each of your points and run OpenCV's keypoint-detector / descriptor then of course you'll most certainly get more keypoints/descriptors.

edit flag offensive delete link more


Thank you, Guanta. I'm sticking with solution 1. It seems to work nicely. I'm now just naivly copying the intitial 2D locations properties to the tracked points.

axjensen gravatar imageaxjensen ( 2013-03-20 02:02:18 -0600 )edit

Question Tools


Asked: 2013-03-19 07:37:53 -0600

Seen: 607 times

Last updated: Mar 19 '13