False robust matching inliers. about opensfm HOT 3 OPEN

Wow, what a detailed analysis! We definitely have some "flying" points in all reconstructions, and this could be one of the main reasons. Super nice find!

I don't see right now a simple way of eliminating those, since they seem to be geometrically valid. Some random ideas:

• use only tracks of length>3, but that might reduce the number of tracks too much
• reduce the ransac threshold when matching to minimize the probability of getting them in. @pmoulon has a nice ransac that estimates the threshold on the fly in openMVG. I wonder if that would reduce the number of wrong matches.
• use some kind of spatial regularization (seems tricky)

from opensfm.

I guess there could always be false symmetric matches that are geometrically valid when using RANSAC even if the threshold is estimated on the fly as in openMVG. It could possibly reduce false matches as you mentioned but probably not eliminate them all since that is the nature of the fundamental matrix.

Increasing minimum track length would reduce the number of tracks significally in a small image set like Berlin. Not sure about larger sets but it seems probable that these contain lots of tracks of length two too.

If the false inliers have a negligible effect on the pose of the cameras it is mostly a visual problem during transition when using the meshed reconstruction as described above.

from opensfm.

how about this check mentioned in opnMVG?

openMVG/openMVG#1219 (comment)

If you have 40 lines (matches) going in the same direction, the two random wrong one will be discarded. When matching feature between images, an algorithm will check if "three nearby features" are still relatively "near by" in the next image. Hence can check if it is or not a real match.

Also is it possible to use fundamental matrix to check previously not matched features (failed ratio test but valid matches) and add them back, this could increase the matches significantly.

from opensfm.