How to reduce the size of the sample regions of SIFT descriptor (instead of 16x16)?

Hello,

I am working with SIFT descriptors for a 2D image classification on python(3.7). As the dimension of the SIFT descriptor vector is 128, the recovered information is less discriminating. Therefore, I applied a PCA to reduce the dimension of my vector to 46 dimension. I was able to improve the result a bit, but I am looking to improve it even more by reducing the size of the SIFT descriptor computation matrix to a size smaller than 16 x 16. Do you have any suggestions on how to do this?

Thank you.

this is usually a “bad idea” (as long as you’re using keypoint-matching for the classification)
so, what are you planning to do here ?

you wont be able to change this from python.

that will reduce the length of the descriptors, but you still have an uneven “count” of those, and thus cannot use ml, like SVM, ANN for classification.
instead, what ppl use is clustering. create a dictionary from your features, later train & classify distance / histogram vectors

  • I try to get the descriptors of my image to build a database where each descriptor I assigned a class manually. Then, using this database and a machine learning algorithm (SVM or RF) I will make an automatic classification.

  • So if I can’t reduce the size of the SIFT descriptor computation matrix to a size smaller than 16x16 on python, is there any other way to do it? I found in an articlehttps://www.researchgate.net/publication/320165788_SIFT_descriptor_to_set_landmarks_on_biological_images that they reduced the size of the computation matrix to 9x9 instead of 16x16 but I don’t know how they did it!

you’ll end up reimplementing SIFT from scratch
(and, noone said, it would get better, by reducing the sample size !)

tried & discarded idea from so many ppl who tried before.
a single descriptor is statistically meaningless.

what you’re trying to do sounds like Content-based image retrieval - Wikipedia but how you’re trying to do it sounds like you haven’t done all that much research into it

yeah, mutilating SIFT isn’t gonna get you anywhere, nor is reducing the dimensions of individual feature descriptors.