I am implementing the ORB feature extraction algorithm, and I have my own FAST implementation. As the paper clearly states, there is no measure of "cornerness" in FAST, which is why Harris is used to give it a score. In OpenCV's source code however, they sort of give a score to FAST keypoints, but I don't understand how.
My problem is that I use Harris to evaluate each keypoint found by FAST (as explained in the ORB paper), in order to keep the best corners only. The result is that I obtain super clustered keypoints in the same areas, which is not good for any application...
What is the cheapest or most efficient way to prevent clustering in FAST? I was thinking of splitting my image in a grid, and then keeping the N top keypoints in each cell, but that requires to compute the Harris measure for every keypoint, and that's too expensive for real-time stuff.... Any ideas ?
↧
