1
$\begingroup$

I'm new to SIFT and was surprised by how much the features changed after resizing an image. Here's the images I used for testing:

Original:

lenna

Resized:

lenna-smaller

Cropped:

lenna-cropped

After running SIFT on those 3 images, I was surprised to find out that the original image was more "similar" to the cropped one than the resized one (I tested with both bag of words + cosine distance and the distance algorithm described in the SIFT paper).

Is that normal and expected? What would be good approach if I wanted the original and resized images to score as more similar than the cropped one? Would normalizing image size before applying SIFT help?

Edit: After drawing the SIFT keypoints on each image, I'm now seeing that the resized image seems to have less keypoints in a common than the original and cropped ones.

enter image description here

I am guessing I should probably normalize image dimensions before doing SIFT to avoid this issue.

$\endgroup$
  • 1
    $\begingroup$ How much more similar? I would expect the original might be similar to both. I'd guess SIFT might have some robustness to resizing (thanks to the image pyramid), and should be highly robust to cropping (there should be a perfect match for the keypoints in the regions of the image that survive cropping). It looks like SIFT was able to compute a good alignment to both images, so it looks like it did its job successfully in both cases. SIFT is a method for aligning images, not for measuring similarity. $\endgroup$ – D.W. Apr 11 '17 at 22:39

Your Answer

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

Browse other questions tagged or ask your own question.