Boost logo

Boost :

Subject: Re: [GSOC 2019] [Draft] Proposal by Olzhas Zhumabek - Epipolar geometry
From: stefan (stefan_at_[hidden])
Date: 2019-04-02 18:15:27

Hi Olzhas,

On 2019-04-02 5:13 p.m., Olzhas Zhumabek wrote:
> Hi,
> Since most of the migration of TMP part of the library done, more well
> researched algorithms proposed by Miral Shirah, I thought I could propose
> some new area that the library doesn't have support for yet. I've taken a
> quick glance around the library and haven't found any tools for epipolar
> geometry. From what I understand, this is a field which emphasizes on
> correspondences between two shots of one scene from two polar positions
> with very little difference in time and gathers information based on that.
> It doesn't require any ML model training or anything like that, from what
> I've seen.


> I'm taking a video lecture course on the topic to estimate how much of a
> far cry this is from existent functionality (how much stuff is missing) and
> estimate how complex it will be to implement this.
> The *list of things that I would to implement* are as following:
> - Algorithm for finding fundamental matrix
> - Algorithm for finding epipolar lines between two images
> - Algorithm for image rectification

For completeness, you should also add some feature detection: To find a
fundamental matrix you need to feed in pairs of points. That is, you
need to already have obtained those corresponding points (as features)
in the original images.

Of course you could assume that something else will provide them, but as
right now we don't have the required algorithm, you need to either
implement that step yourself, or hope that someone else will, in time. :-)

> I will add/remove/expand some items as I move on and gather more
> information.
> So I have the following questions:
> Is anybody working on this? I quickly sifted through git branches and
> haven't found anything related to this.

I'm not aware of anyone working on this.

> Is this outside the scope of the library? I believe as long as it doesn't
> involve machine learning or some other training stuff, it should be ok.

Machine learning isn't the only thing that doesn't fit the scope of GIL. :-)

More seriously, I think this would be a nice application of the library.
I'm not entirely sure that all of the algorithms fit *into* the scope of
GIL, but at least, this would be a wonderful application of GIL, and so,
I'd be very happy to see some contributions in that area.

> Is there anybody who has some experience implementing this kind of
> algorithms?

I personally don't, but I'm talking to some colleagues of mine who have...

> I will modify the proposal to include some more concrete implementation
> details/architecture after delving deeper into the topic.

Great, thanks.


       ...ich hab' noch einen Koffer in Berlin...

Boost list run by Boost-Gil-Owners