Find similar images fast

asked 2013-12-12 02:42:21 -0500

a19a gravatar image

Hi all, I have a dataset of 200,000 images and I need to build a fast CBIR system. I came up with OpenCV, but I am totally newbie to the argument.

From what I understand I could create a FLANN index of SURF (or BRISK?) features, and then perform a KNN search over it.

I didn't find any step-by-step tutorial or example on the topic, so I really don't know where to start.

Could you help me?


edit retag flag offensive close merge delete


Ist the comparison a 1:1 comparison or are there any "transformations" between your images such as:

  • affine transforms, rotation, scale, translation
  • perspective transformations such as tilt etc.
  • noise
  • changes in luminance, contrast etc.

... or a combination of these transforms? If there are affine or perspective transforms, then feature matching is your friend. If not, try to use a method like cross-correlation, template matching etc.

JohannesZ gravatar imageJohannesZ ( 2013-12-12 04:49:07 -0500 )edit

Thanks for replying. I don't I told I am a totally newbie to OpenCV. All that I know is that I need a CBIR fast retrieval system with a dataset of 200,000 images. I would need a step- by- step tutorial.

a19a gravatar imagea19a ( 2013-12-12 05:00:23 -0500 )edit

could you describe a bit, what 'CBIR' means exactly to you here ?

do you just need 'some similarity score' or have it sorted into one ore more 'categories' ? (mountain,pet,tool)

berak gravatar imageberak ( 2013-12-12 13:26:41 -0500 )edit

Also, can you upload some example images?

GilLevi gravatar imageGilLevi ( 2013-12-15 04:14:21 -0500 )edit