It may be hard to see, but these are the correspondences found by the Okutomi algorithm. They're not 100% accurate, but its pretty close, and it can be improved by expanding the template-matching window size.
Unfortunately, the algorithm is pretty slow. I can find these 100 correspondences in between 5-40 seconds (depending on locality of the points), so finding a dense mapping (~800*800 points) is infeasible. I've tried all optimizations short of compiler options (which never give the light-and-day differences you'd hope for). I'm considering shrinking the image.
The biggest challenges ahead are the transformation and the performance. These problems are pretty tightly related, so I'll start with the transformation. A homography does terribly (obviously...), but I picked up some literature on the subject at the library today: http://www.imgfsr.com/ifsr_book.html, and downloaded some sample code: http://biocomp.cnb.csic.es/~iarganda/bUnwarpJ/ (two camera view and in java). I'm currently wrapping my head around non-rigid transformations so I can weigh my options and decide where to go next.
No comments:
Post a Comment