[Bf-committers] LibMV versus OpenCV / VXL

[Troy Sobotka]

On Thu, Apr 21, 2011 at 2:27 AM, Matthias Fauconneau
<matthias.fauconneau at gmail.com> wrote:
> While OpenCV or VXL are only bricks which may help you implement match
> moving algorithms, the goal of libmv is to implement a complete
> tracking system in the library so that an application (like Blender)
> can use it without any algorithmic knowledge.

I am aware of the design goals of libmv. My issue is that:

1) Computer vision is obviously complicated. Relying on an internal
representation library may open Blender up to significant additional
bugs and further tax developers. Judging from the libmv stagnation, I
think this isn't too much of a stretch of the imagination.
2) The existing libraries of VXL and OpenCV have the vast majority of
the hard work already done, as well as a set of interested eyes and
developers monitoring the code base.
3) Fresh technology. OpenCV implemented SIFT recently thanks to a
relatively healthy developer community interested in computing vision.
This would also benefit Blender with a limited amount of resources to
update the code base.

> It is now being rewritten as a complete modular framework with clean
> interfaces and automatic testing.
> I think improving libmv and using it in Blender is the best way to
> implement the first open source match moving application.

And I don't disagree entirely.

I am just curious if perhaps the upsides versus downsides have been
fully considered at this point.

Judging from the code sample for optical flow tracking, I wonder how
much more difficult it would be to see a native implementation in
Blender based off of one of those two libraries.

With respect,
TJS