SIFT+RANSAC matching algorithm

asked 2017-02-19 08:58:58 -0500

Machupicchu gravatar image

updated 2017-02-19 09:23:06 -0500


I ve wirtten some code to compute SIFT decriptors from images and then compute the homography matrix from the 'good matches'. My goal is to distinguish between 2 classes (different and same) images.

[H, mask ]= findHomography(...)

However the values returned in mask are almost the same (number of inliers between 4-8) for both 'same' and 'different'.

My procedure was to check the number of inliers in mask ... but is that the correct way? Or should I apply the Homography to my image and rerun RANSAC and retest something like the so-called 'back-projection error'? If yes can you provide some hints about the threshold to use?

Best regards. O.

edit retag flag offensive close merge delete


My goal is to distinguish between 2 classes

then you got the wrong algorithm for this.

feature matching is meant to produce a homography for a known scene between 2 images, it's not meant to distinguish between 2 different ones. the outliers in the homography only specify, which points were acceptable for the transformation, they do not measure similarity at all.

berak gravatar imageberak ( 2017-02-20 03:21:57 -0500 )edit

I didn't explain clearly maybe: the distinction is intended to be made by the percentage of error with the homography , i.e. if too much error (threshold) then no match -> not able to realign properly, otherwsise match

Machupicchu gravatar imageMachupicchu ( 2017-02-20 10:12:53 -0500 )edit

still, wrong tool for your job. none of those measurements give you a meaningful similarity.

berak gravatar imageberak ( 2017-02-20 10:47:55 -0500 )edit

I guess I am still not formulating it well enough. I can assure you it is used in academics domain (university) to to such kind of evaluation (detection), therefore I am sure this approach is valid. The idea is that we use the geometrical info in the keypoints to try realingning them to a supposedly matching image. If alignment is good->match if no good-> no match maybe you see better what i mean now ?

Machupicchu gravatar imageMachupicchu ( 2017-02-20 19:31:52 -0500 )edit