Precision of this exercise

[JanEisklar]

Hi,

I just wanted to ask how good the results of this exercise should be.

Is this the desired output or is there something wrong as there are still some glitches remaining?


And here are the interest point matches from ransac:


Thank you!

[Christoph Vogel]

Your result looks almost correct.

But it seems that you do not find the best homography.
If you visualize your result without blending, that should be clear.

Your homography points are very localized, therefore small errors in the estimation might become more severe.


Best regards,
Christoph

[JanEisklar]

Hi,

we modified our ransac implementation so that it not only tries to get the most inliers, but also maximizes the distance between the points.
Unfortunately though, the results are still the same:





Any ideas?
Are we the only ones to experience this problem?

EDIT:

Oh btw. here is the output of displaymatches for the images above.

Code: Alles auswählen

Found 20 inliers!
Match 1: dist=0.27914
Match 2: dist=0.47025
Match 3: dist=0.71193
Match 4: dist=0.87769
Match 5: dist=0.89023
Match 6: dist=0.98987
Match 7: dist=1.0719
Match 8: dist=1.1897
Match 9: dist=1.3035
Match 10: dist=1.3532
Match 11: dist=1.4211
Match 12: dist=1.5414
Match 13: dist=1.5612
Match 14: dist=1.5863
Match 15: dist=1.74
Match 16: dist=1.806
Match 17: dist=1.8194
Match 18: dist=1.8287
Match 19: dist=1.8701
Match 20: dist=1.9939

Distances of <2 shouldn't cause such a bad homography, so it must be something else we are missing

[Christoph Vogel]

Hello.

A distance of 2 is fine, but I think there should be a lot more inliers.

So either you have too few inliers from the start or you do not find all of them.

When visualizing your best pairs, how many inliers do you count ?

With my parameter setting I have a rather large fraction of inliers.
I get that result with the settings provided (window size 50).
You can also compare your result with the pictures from the slides presented at the exercise last week.

If you count less inliers your problem should be in the first part of the exercise.
Otherwise your ransac (exiting too early?) or homography estimation has problems.

[JanEisklar]

Our implementation of the homography estimation was erroneous.
It's amazing how "good" our results were though, given the severity of this mistake.

Anyway..thank you.