There are 24 matches with 2 outliers (#22 and #23 at the bottom). Lowering the threshold results in even more matches (with more outliers as well), but RANSAC should be able to work with only a handful of matches.
Saturday, February 3, 2007
Feature Correspondence Working
I followed Serge's recommendation to use Ben-Haim Features -- the difference between normal NCC is you use the X and Y components of the image gradient instead of the raw image. The two scores are then added together for the total score. I'm not sure why, but the results are surprisingly accurate. Here's what it looks like applied to my globe model with a threshold of 0.94:
There are 24 matches with 2 outliers (#22 and #23 at the bottom). Lowering the threshold results in even more matches (with more outliers as well), but RANSAC should be able to work with only a handful of matches.
There are 24 matches with 2 outliers (#22 and #23 at the bottom). Lowering the threshold results in even more matches (with more outliers as well), but RANSAC should be able to work with only a handful of matches.
Subscribe to:
Post Comments (Atom)
1 comment:
Hi. Interesting post you've made.
Though I don't really understand what is the main goal of your project, but it seems interesting :).
Anyway, if you are under time constraints you should think twice, before implementing sift. Even if you are really really clever and good programmer, my estimation, it will take you at least one month just to implement and test algorithm - partially, because algorithm is not clear in some places. My advice: don't spend time, unless you really need to. Good luck.
Post a Comment