Re: Five Camera Stitch debate
- Hi everyone,
To be honest, we stopped using a 5 camera cluster a long time ago, in favor
of mirror rendering in Blender. The calculations I shared got the best
results for us, but we may have been overshooting the mark.
With testing, you should be able to optimize the source renders. I wouldn't
be surprised if your optimized number is something lower than what I
suggested but higher than a through-the-zenith count.
It seems a shame that 40% of the rendered pixels are thrown out to make the
fisheye. Masking the lower half of the surrounding cameras is a great
Paul's idea is good too -- there is also a three camera scheme floating
In our end: since switching to mirror rendering in Blender, our average
render time for a 2000x2000 fisheye is about 20 seconds per frame per CPU
(depending on complexity). The interesting thing is that a regular camera
takes a longer time to render at the same resolution (in Blender).
You can check out our camera set up here:
Ott Planetarium - Weber State University
- There are times when one approach work better than others...
sometimes there is no choice... and it's all software dependent.
Here at Home Run Pictures, we use several different approaches... a
multi-camera cluster, a single camera view, etc. Depending on what
we are rendering determines which path we take... either because of
faster rendering, what rendering CPUs are available, not needing a
full circle render (small element in the scene), or just because a
certain camera approach will not render an effect correctly. At
times a single camera approach is very memory intensive and the multi-
camera cluster is preferable since you are only seeing a smaller area
of the scene with each camera. Other times a refracting or
reflecting path will not allow certain effects to render correctly,
like glows, reflections or shadows, etc.
It all depends on what your end need is. As they say, "the end
justifies the means," at least in rendering fulldome scenes, which
always is a challenge... so pick your poison.
On Oct 16, 2008, at 1:55 PM, Ron Proctor wrote:
To be honest, we stopped using a 5 camera cluster a long time ago,
in favor of mirror rendering in Blender.
H o m e R u n P i c t u r e s
President & Creative Director
100 First Avenue - Suite 450
Pittsburgh, PA 15222
> At times a single camera approach is very memory intensive and the multi-Ahhh, I remember it well ... doing the Mars (MOLA) flyover for "Infinity
> camera cluster is preferable since you are only seeing a smaller area
> of the scene with each camera.
Express" for SkySkan based upon the (then exciting) 1/32 data of the whole
planet. I had the choice of fisheye or cubic maps but it was indeed better
to use cubic maps because it allowed more pruning of the triangle database
than what one could do with rendering fisheye directly. Hence the original
reason for writing my cubic map to fisheye stitcher.
A bit of history for those who care about such things.