|
I have used SIFT to find keypoints and descriptors,now I want to decrease number of descriptors because for example for an image with size 256x256 it gives me 120000 descriptors which is time consuming. please let me know any suggestion how to decrease number of descriptors. Thanks |
|
One dimensionality reduction approach -- assume that you are looking for k-prototypical descriptors. That is you define how many descriptors you want to have - and let that be k. Then cluster the descriptors into k-clusters. Pick the centroids of each cluster. You now have k-most representative descriptors. Now every descriptor can be expressed in relation to these k-prototypes to get a k-dimensional vector. The i-th component of the vector can be the Euclidean distance (or any other distance/similarity measure) between the descriptor and the i-th prototypical vector. Thanks for your answer, if i reduce the dimension of the image some of image information will be lost. my main task is to find similar areas in one image for example if we copy some part of the image and paste it to another part so these two parts are similar, i have used Euclidean distance after applying SIFT on tampered image. how I can define how many descriptor I want ? and how should I cluster the descriptors into k-clusters?
(Nov 08 '12 at 01:00)
Laleh
|
|
It's depends on the kind of application you are looking for. In case you want to do SIFT matching (David Lowe's first SIFT paper) then it's advisable not to cluster keypoints. If you want to do classification (holistic like objects) then you can still cluster points. You can also look at work from Nister and Stewenius where they do matching by clustering keypoints. [Deleted what looked to be a duplicate post.]
(Nov 04 '12 at 01:24)
gdahl ♦
@karan sikka: thanks for your answer, I am using David Lowe's SIFT to find similar descriptors in one image, if I don't use cluster then what should I use? what is your suggestion?
(Nov 06 '12 at 23:41)
Laleh
@karan sikka:please let me know where I can find Nister and Stewenius works?
(Nov 08 '12 at 03:31)
Laleh
|
consider just keeping the descriptors otherwise look at dimensonality reduction with things like pca or autoencoders
thanks for your answer