Ask Your Question
0

Does SIFT work the same on different images sizes?

asked 2017-06-02 06:23:57 -0600

swiss_knight gravatar image

I was wondering if the result of any keypoint detector & descriptor algorithm is "input-image-size" dependent.

For example, with the SIFT algoritm, I don't have the same kp and descriptor (some are, but lots are not) when working on full size sized images or (down-)resized images (original: 1000px, resized: 500px e.g.).

On bad quality pictures (foggy landscapes e.g.), it may be an issue when working on reduced size images (in order to improve computation time when dealing with thousands of input images).

Is there a way to improve that?

edit retag flag offensive close merge delete

2 answers

Sort by ยป oldest newest most voted
0

answered 2017-06-14 11:33:32 -0600

swiss_knight gravatar image

Thanks.
Yes SIFT is scale invariant by definition for a given image. But between two images, let say one 400px width and its 3200px width original the keypoints founds are absolutely not the same.

For example, on the big image, >90% of the retained keypoints for the matches between this image and another large one, are on a wall in the background, all located in the top left corner. But on the small images, all matches are spread a little everywhere on the image, there is no more keypoints matches on the wall.
Thus, kp and matches are strongly dependent of the original image size.

edit flag offensive delete link more
0

answered 2017-06-02 07:03:13 -0600

updated 2017-06-02 07:04:40 -0600

You want it to work the same on all images, that is how it attains scale invariance (though it does adjust number of octaves given the image resolution). Other things you could try include not using the default parameters and instead trying to find parameters more suited to your use case. However, the best algorithms and methods are those that you don't need to adjust parameters to achieve useable results. The default SIFT parameters are probably chosen with general applicability in mind. If you haven't read the original SIFT paper, I suggest doing so, it is easy to digest.

edit flag offensive delete link more

Question Tools

1 follower

Stats

Asked: 2017-06-02 06:23:57 -0600

Seen: 2,184 times

Last updated: Jun 14 '17