flann feature matching - image distance

asked 2015-05-06 06:01:38 -0600

I have compiled and implemented this tutorial: http://docs.opencv.org/doc/tutorials/...

My objective has been to compare a number of pairs of images to identify which ones are the same image (Identical image but can be slightly modified - cropped, brightness, text ontop).

I thought that if I gathered the number of "good matches" when comparing 2 images this might give me a measure on how similar they are. Im finding that this is not the case, I have non matches with 400 good match references and matches that have 15 good match references.

I guess this may depend on the image content as to how many feature points are found.

Im keen to see how SURF/SIFT can identify identical images, Any advice on how to get an output of "Match" or "Non match"?

Any distance number that i can tune with my image set would be ideal

answered 2015-05-06 07:35:44 -0600

JohannesZ
419 ●1 ●2 ●14

First, the number of points depends always on your texture. SIFT/SURF extract their features in corner-like image structures. So no corners, no/less feature points of poor quality.

In general, there is no hard border for match / non match. You may set a specific threshold but this wilk not give you robust results. A solution is to work with probabilities. Imagine a histogramm which tells you "model A gets 20 matches", "model B gets 2 matches" and "model C gets 3 matches". In this case, your distribution will tell you what will be a good match. Also try to look for terms like "bag of words classifier", "k-means" etc. OpenCV will help you with this.

edit flag offensive delete link

Comments

Thanks for the answer. Number of feature points would definitely vary per image so its not a good judge of match success.

I was hoping it would be possible to gather a distance type metric between each of the feature points on each image. Ie. If there are 20 points detected on one image and then 19 of the same points are detected on another image (within x "distance") this would be a strong match?

Bag of visual words, k-means etc align around training against a large image set. Im looking at one image compared to another image. I was hoping this would be possible as a one off comparison?

mattnthat ( 2015-05-11 05:55:42 -0600 )edit

add a comment

flann feature matching - image distance

1 answer

Comments

Links

Question Tools

Stats

Related questions

flann feature matching - image distance edit

1 answer

Comments

Links

Question Tools

Stats

Related questions

flann feature matching - image distance