CODEWITHSUNDEEP
opencvfeature-detectionfeature-extractionfeature-selectionoutliers
Michele mpp Marostica
Michele mpp Marostica

Reputation: 2472

OpenCV: How to compare the performances of feature detectors and extractors using images of a real scene by different point of view?

I'm trying to build a loop closure algorithm, but before to start the development I would like to test which feature descriptor work better on a real dataset.

I have a pair of images of a corridor taken in both direction, one entering in a room and the other leaving the same room. So they represent the same scene but from 2 different (opposite) points of view.

I'm using OpenCV and I've already coded the part for the detector, the extractor and the matcher.

But now looking at the results I find that there are a lot of false matches. How can I drop them? I tried to use cv::FindFoundamentalMat but I'm not sure that this is the right way to do that.

edit

I found this paper (ShmidtJAMRIS12) that exactly does what I have in mind but I cannot understand how they choose the parameters of each detector and extractor

Upvotes: 0

Views: 308

Answers (1)

ChronoTrigger
ChronoTrigger

Reputation: 8617

Finding some kind of geometrical consistency is one of the basic ways of ensuring two matching images are the same place. A fundamental matrix may be ok, a homography may be ok if you match planar surfaces (as walls in corridors). If you work with a stream of images, you can also check if matches of consecutive images are consistent. Here is a paper that uses these techniques: http://webdiis.unizar.es/~dorian/dl.php?dlp=GalvezTRO12.pdf

Upvotes: 1

Related Questions