Estimating iterations and match quality

Moderator: Computer Vision

Flix
Erstie
Erstie
Beiträge: 20
Registriert: 14. Jan 2013 12:37

Estimating iterations and match quality

Beitrag von Flix » 19. Jul 2013 12:08

Hi,

I don't get how the number of iterations necessary for RANSAC is calculated. The formula from l10 slide 26 is pretty simple of course, but how do we get the probability of an inlier? Do we have to guess it completely random? Just to be sure: inliers in our case are point pairs that are matched correctly, i.e. both points in the pair represent the same point in world space?!

As I'm not able to get any decent homographies at all, could someone please have a look at my input and tell me, whether the matches I work with are accurate enough? I filtered out point pairs above a certain chi-squared threshold, but I'm not sure if this is sufficient, there are still a lot of nasty diagonals ...
asdf.png
asdf.png (171.54 KiB) 891 mal betrachtet
Thanks a lot!

lustiz
Mausschubser
Mausschubser
Beiträge: 70
Registriert: 29. Apr 2009 10:28

Re: Estimating iterations and match quality

Beitrag von lustiz » 19. Jul 2013 12:42

Flix hat geschrieben:The formula from l10 slide 26 is pretty simple of course, but how do we get the probability of an inlier? Do we have to guess it completely random?
That's right. Just take a guess. It doesn't matter that much though, after all it's just about computing a reasonable number of iterations that is good enough but not too big. The task says we should make a conservative guess, how many percentage of inliers do you think you have at most? If you figured that out just subtract a couple of percentages :P
Flix hat geschrieben: Just to be sure: inliers in our case are point pairs that are matched correctly, i.e. both points in the pair represent the same point in world space?!
Yeah, it basically means, you have connected the exact same spots in the left and the right image through your lines.
Flix hat geschrieben: As I'm not able to get any decent homographies at all, could someone please have a look at my input and tell me, whether the matches I work with are accurate enough?
It kinda looks fine, but I have a decent couple of more pairs?
Flix hat geschrieben:I filtered out point pairs above a certain chi-squared threshold, but I'm not sure if this is sufficient, there are still a lot of nasty diagonals ...
Why? You do not need to do that! "Diagonals" are nothing to worry about. RANSAC will judge if your pairs are inliers or not, but that is not up to your thresholding beforehand.

Antworten

Zurück zu „Computer Vision“