Feature Detection

From: Olzhas Zhumabek (anonymous.from.applecity_at_[hidden])
Date: 2019-06-07 04:04:58

Hi,

So I have bad news, sort of good news, and some more info.

Lets get bad news out of the way first: ORB's feature descriptor needs
learning to produce better results. So while implementing it, I probably
will not be able to tell for sure if it is bug in my code or that the
descriptor is simply not capable of doing any better.

Now onto sort of a good news: most of the founding blocks for ORB are
ready. The rest is basically calling functions. I believe it shouldn't take
long to implement it, I would guess 2-3 days, if I understood the paper
correctly.

Info:

I found one author on github and posted an issue if I can make another
implementation: https://github.com/pablofdezalc/akaze/issues/36 . One of
the authors responded by being okay in the original email I sent to them,
and that author is second one. There are three authors of KAZE, and I
believe different authors for accelerated-KAZE.
The latter seems to be easier to implement, so setback from time spent on
lanczos shouldn't affect too much.

I also found emails of authors of FREAK, the predecessor to RIFF. From what
I understood, FREAK can be used to obtain keypoint information of a
relatively large patch of image, and then used to find it in another image.
I believe we could somehow combine less accurate keypoint detector with
this descriptor to improve accuracy. Not sure how effective it will be
though. I sent an email to them if it is okay by them for me to implement
it.

Extensive patent search for RIFF yielded nothing either. It would be great
if there would be a way to try to implement it and quickly takedown if
there is a patent, but I guess that is impossible.

I can leave original feature descriptor that ORB modified, I believe it is
called BRIEF. It will not be invariant to rotation, e.g. rotating the image
will be detrimental to matching.

Olzhas