2.3 microsecond feature point matching

From: http://www.cvchina.info/2010/02/10/2-3-us-feature-point-match/

 

Demo Video

 

This is cvpr2009
The best demo. From Cambridge University. Match a feature
Point is only 2.3 microseconds. It is microsecond, not millisecond! A plane image is calculated based on 200 feature points (200 points mentioned in the paper are enough to estimate the posture), So 2000 frames can be processed in one second !! Trenge
God. From the above video, thisAlgorithmIt can run smoothly on the mobile phone. It's awesome. Can you imagine using sift or surf on your phone...

The algorithm logic follows ferns. However, fast9 is used for feature point detection, and features become binary features. sse2 can be used for acceleration or dynamic matching, not very attractive. Let's look at the paper.
Right. We hope this algorithm can be added to the next version of opencv.

The project homepage is here
. You can watch the large image video and PPT.