# Matcher distance and neighbours… how?

I wonder how distance and neighbours work together. Do distance set a limit for neighbours?

1 Like

A question I asked in another thread last week, Saijin was going to look into it.

I would have thought just setting a distance would define the number of neighbours, but maybe the number limits it to the closest n images within the distance?

And I wonder why both numbers is needed. I think distance would be enuf really if you got gps data.

1 Like

And if there is no GPS data, how does it know which images are the neighbours?

1 Like

It doesn’t in that case. You’ll notice if there’s no gps, it switches to BOW matching, which you can think of as using a sorted index of features to figure out which images are most likely to be matched.

2 Likes

That sounds very time consuming.

Any ideas on how the matching task is approached (with GPS), does it try closest first, or something else? Just watching my current big task, which has been going 140 hours now, it does seem a bit random

2021-11-23 15:26:24,700 DEBUG: Matching DJI_0868_9.JPG and DJI_0867_9.JPG. Matcher: FLANN (symmetric) T-desc: 91.191 T-robust: 0.101 T-total: 91.333 Matches: 5098 Robust: 5051 Success: True
2021-11-23 15:26:24,718 DEBUG: No segmentation for DJI_0779_10.JPG, no features masked.
2021-11-23 15:26:25,766 DEBUG: No segmentation for DJI_0322_7.JPG, no features masked.
2021-11-23 15:26:26,005 DEBUG: Matching DJI_0772_3.JPG and DJI_0743_3.JPG. Matcher: FLANN (symmetric) T-desc: 56.012 T-robust: 0.017 T-total: 56.040 Matches: 26 Robust: 9 Success: False
2021-11-23 15:26:27,880 DEBUG: No segmentation for DJI_0315_7.JPG, no features masked.
2021-11-23 15:26:28,642 DEBUG: No segmentation for DJI_0127_8.JPG, no features masked.
2021-11-23 15:26:31,043 DEBUG: Matching DJI_0322.JPG and DJI_0317.JPG. Matcher: FLANN (symmetric) T-desc: 62.153 T-robust: 0.001 T-total: 62.158 Matches: 107 Robust: 82 Success: True
2021-11-23 15:26:31,227 DEBUG: Matching DJI_0927_6.JPG and DJI_0924_6.JPG. Matcher: FLANN (symmetric) T-desc: 82.943 T-robust: 0.028 T-total: 82.987 Matches: 2008 Robust: 1981 Success: True
2021-11-23 15:26:33,606 DEBUG: No segmentation for DJI_0678_6.JPG, no features masked.
2021-11-23 15:26:33,771 DEBUG: No segmentation for DJI_0083_4.JPG, no features masked.
2021-11-23 15:26:34,412 DEBUG: Matching DJI_0583.JPG and DJI_0590.JPG. Matcher: FLANN (symmetric) T-desc: 66.070 T-robust: 0.016 T-total: 66.104 Matches: 231 Robust: 208 Success: True
2021-11-23 15:26:35,625 DEBUG: No segmentation for DJI_0520_9.JPG, no features masked.
2021-11-23 15:26:37,025 DEBUG: No segmentation for DJI_0092_4.JPG, no features masked.
2021-11-23 15:26:37,179 DEBUG: No segmentation for DJI_0382.JPG, no features masked.
2021-11-23 15:26:38,098 DEBUG: Matching DJI_0237_5.JPG and DJI_0709_6.JPG. Matcher: FLANN (symmetric) T-desc: 65.969 T-robust: 0.018 T-total: 66.011 Matches: 32 Robust: 12 Success: False

I’ve been watching, and it appears that a successful match requires at least 20 robust matched features.

1 Like

Yes and no. It isn’t exponentially expensive computationally due to the indexing, but I don’t know what the cost is (haven’t read the papers…).

2 Likes

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.