[Draft] Proposal by Olzhas Zhumabek - Epipolar geometry

Subject: [Draft] Proposal by Olzhas Zhumabek - Epipolar geometry
From: Olzhas Zhumabek (anonymous.from.applecity_at_[hidden])
Date: 2019-04-06 22:29:43

Hi,

I've prepared a draft
<https://docs.google.com/document/d/1cfj3mycgW1ItVO1DpsZiQRpeisUM0fvE3HszOV1VS5M/edit?usp=sharing>.
It is very rough, I will add technical details about KAZE feature detector
and some other parts. Could you please have a look at it? There was a share
button, but it had so much warning text on it I decided not to press it yet.

About feature detector: it seems like both SIFT and SURF are patented, and
patents will not expire in this year. So I opted into KAZE, but I'm not
sure if it is patented. I've sent an email to authors to verify if there
are no patents that would infringe their copyright. The detector seems to
be both comparably accurate to SIFT and comparably fast to SURF.

 reconstruction (mapping 2D features to 3D features)

I'm not planning to support 3D directly at this point. Fundamental matrix
calculation does not require knowledge about relative positions of cameras,
but of course the matrix will have little correspondence to real world.
IIRC there can be many fundamental matrices.

> Doesn't this depend a lot on context ? How far are the two cameras apart
> ? If they are close enough together, one might start the search in image
> 2 using the positions in image 1 as a starting point. If not, that may
> not be a good strategy.

Feature descriptor should possess enough information to be uniquely
identifiable and reproducible in the other image. I believe that most of
the natural scenes will not trigger false positives, but some edited photo,
like copy pasting the same object and lining it up will certainly pose a
problem.

Best,
Olzhas