0
$\begingroup$

Is there a better way to increase Recall Rate when using SIFT features?
I am thinking a way to replace the NN1/NN2 ratio to account for slightly distorted objects.
Moving towards clustering and using BOW(Bag Of Words) seems a way but I need to do one-on-one match of objects in images rather than training and learning. This refrains me from thinking towards BOW.

Anybody got any idea?

|improve this question
$\endgroup$
1
$\begingroup$

With slight modification you might want to use RootSift:

http://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf

Also the other steps in the paper will guide to improve the recall rate.

Cheers,

|improve this answer
$\endgroup$
0
$\begingroup$

You can just increase the threshold for NN1/NN2 ratio. By the way, which detector do you use? In "Two-view Matching with View Synthesis Revisited" paper have been shown that different detectors need slightly different ratio. Also they recommend to take into account geometrical position of the NN2 in order to increase ratio.

But it is always balance between recall and precision - you increase recall and loose precision. Could you explain what exactly do you want to do - which images and what the final result you want to archive?

|improve this answer
$\endgroup$
  • $\begingroup$ I am using SIFT detectors and SIFT descriptors NN1/NN2 ratio based matching But NN1/NN2 gives a large number of false matches too and they don't work on distorted images. I want to get more matches in first place and thereafter filter out incorrect matches. But it seems NN1/NN2 removes possible candidate for match. This might be due to : 1. Noise matching more precisely 2. 2-3 points lie around each other among which 1 is correct match causes NN1/NN2 condition to fail 3. The actual correct match(due to small distortion) may be farther than the nearest match(noise) $\endgroup$ – object recognition Feb 16 '14 at 10:50
  • $\begingroup$ NN1/NN2 works on assumption that the nearest one is the correct match(if at all) and then finds the match confidence by the ratio. I tried 3-NN match for all keypoints but they lead to enormous number of matches which is very difficult to filter afterwards. $\endgroup$ – object recognition Feb 16 '14 at 10:51
  • $\begingroup$ I think that in BOW, clustering allows for small distortion and hence leads to robustness. False visual word presence can be removed by learning from images which words are important for that object and hence weight it ("tf-idf weighting" as the term says). But I need to match object one-on-one,i.e., I need to find presence of a given object in images (most of which are false). Can I use BOW model for this purpose? If yes, a better insight of how to do it is highly appreciated. $\endgroup$ – object recognition Feb 16 '14 at 10:52
  • $\begingroup$ What do you mean exactly by "distorted images"? Could you attach some examples of the images you try to match? > 2. 2-3 points lie around each other among which 1 is correct match causes NN1/NN2 condition to fail This problem is solved in the paper I have mentioned above. > But I need to match object one-on-one,i.e., I need to find presence of a given object in images (most of which are false). May be you need something like Viola-Jones approach - to train some classifier? Again - I cannot advice much without understanding which kind of images you have. $\endgroup$ – old-ufo Feb 16 '14 at 11:54
  • $\begingroup$ BTW, you can try binaries from that paper - cmp.felk.cvut.cz/wbs $\endgroup$ – old-ufo Feb 16 '14 at 11:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.