Implement SIFT feature detector and descriptor #5472
Conversation
…nto feature/sift
…oints in one array
There was a problem hiding this comment.
I think this is in good shape.
We probably do want a commit to remove _oversample_bilin, but we can preserve it in the commit history. Lately, we typically squash all commits, but this seems like a case where it could be worth doing a manual rebase to squash a lot of the smaller ones, but preserve some important changes/revisions. I am happy to help with that rebase once we have another approval here.
|
Fixes #4499 |
|
|
||
| This example demonstrates the SIFT feature detection and its description | ||
| algorithm. | ||
|
|
There was a problem hiding this comment.
It would be nice to have some context here about what SIFT is, how it is used, and what other feature detectors exist in skimage.
There was a problem hiding this comment.
Sorry for the delay, I was quite busy :D How about:
This example demonstrates the SIFT feature detection and its description
algorithm.
Even though it was already published in 1999, SIFT is still one of the most
popular feature detectors available, as its promises to be "invariant to image
scaling, translation, and rotation, and partially in-variant to illumination
changes and affine or 3D projection" [1]. Its biggest drawback is its runtime,
that's said to be "at two orders of magnitude" [2] slower than ORB, which makes
it unsuitable for real-time applications.
.. [1] D.G. Lowe. "Object recognition from local scale-invariant
features", Proceedings of the Seventh IEEE International
Conference on Computer Vision, 1999, vol.2, pp. 1150-1157.
:DOI:10.1109/ICCV.1999.790410
.. [2] Ethan Rublee, Vincent Rabaud, Kurt Konolige and Gary Bradski
"ORB: An efficient alternative to SIFT and SURF"
http://www.vision.cs.chubu.ac.jp/CV-R/pdf/Rublee_iccv2011.pdf
There was a problem hiding this comment.
It looks pretty good to me, but I would modify slightly to spell out SIFT early on:
"The scale-invariant feature transform (SIFT) was published in 1999 and is still one of the most popular..."
There was a problem hiding this comment.
Perhaps a reference to the wikipedia article could also help for those wanting a bit more detail in a less technical format than the conference publications:
https://en.wikipedia.org/wiki/Scale-invariant_feature_transform
There was a problem hiding this comment.
I noticed that the link to the paper "ORB: An efficient alternative to SIFT and SURF" does not work anymore. I hope it's alright, that I changed it. Should we also change it in orb.py?
…nto feature/sift
| return deltas | ||
|
|
||
| def _set_number_of_octaves(self, image_shape): | ||
| size_min = 12 # minimum size of last octave |
There was a problem hiding this comment.
Shouldn't size_min be a parameter here?
There was a problem hiding this comment.
I thought about that too. In the original implementation this is a constant and can't be changed. My reason not to do it was that the user can still practically increase size_min by lowering n_octaves and I'm not sure if it makes sense to lower size_min even more.
But if you prefer it as a parameter I have no problem with that. We just need to find some minimal value, because size_min too low would probably start to break some things.
There was a problem hiding this comment.
An other option would be to make it private attribute of the SIFT class to make it twickable for "expert" users...
There was a problem hiding this comment.
That sounds good, I will add it.
There was a problem hiding this comment.
I plan to merge this PR soon for inclusion in the 0.19 release. If this does not get addressed before then, it can go in a follow-up PR as it would not be a public API change.
There was a problem hiding this comment.
Let's leave it as is for now. We can always add this feature if it is requested, but if it is not a parameter in the original it seems unlikely. Also, I am not enthusiastic about private attributes tweakable by expert users: who should either have an API for it, or it should not be adjustable.
|
Okay, this is now in. Thank you, @faberno for the large amount of work and your patience seeing this through! |
|
Thanks for all of your support! It was a lot of fun and in the future I'll be happy to contribute again. |
|
It is fantastic to see SIFT finally implemented in scikit-image. Thanks very much @faberno for this great contribution! |
Description
Fixes #4499
This PR implements the SIFT feature detection and its descriptor. It was originally developed by David Lowe (https://doi.org/10.1109/ICCV.1999.790410), but this implementation is is mostly based on the paper "Anatomy of the SIFT Method" and its demo by Rey Otero (https://doi.org/10.5201/ipol.2014.82).
I tried to mostly copy the style of ORB with the three Methods: detect, extract, detect_and_extract.
Basic tests and an example in ./doc/examples are provided.
For reviewers
later.
__init__.py.doc/release/release_dev.rst.