My 2nd attempt to apply the SURF feature extraction (vision.ee.ethz.ch/~surf/) functionality of recent OpenCV versions on video content. Finally got it to work when wrapped in OF, after running in some weird memory leak using only standard C++. The original demo code is largely based on this one for static images by Liu Liu: is.gd/NAw7 (The image is by me for my little son)
I've been very impressed to see how well the image and its bounding quad are still recognized even if quite a large parts are occluded...
Next steps are to wrap this all in a nice new class for the existing ofxOpenCV addon, add temporal smoothing to stabilize the found marker edges/corners, make it threadsafe and add some additional matrix wrangling for using this all in markerless augmented reality scenarios (our real target here...)
UPDATE: by tweaking the SURF threshold parameter I could improve the framerate to be around 10fps whilst retaining fairly stable tracking.