Re: [Openexr-devel] Stereo (multi-view) pipeline

[Peter Hillman]

Hi Ton,

There was some discussion about the positioning of the 'views' name part 
when developing the multiview EXR standard. As you say, there's little 
difference from a developer point of view.

It seemed to us nicer that an alphabetical list of channels shows layers 
and passes first. It's easier to navigate that channel list to select 
individual passes. Arguably, the fact there are multiple views of a pass 
in an EXR is only as interesting to the artists as that it contains 
R,G,B,A channels. Conceptually, the image contains "stereo layers and 
passes" rather than two separate sets of layers and passes, one for each 
view. Now that the standard is more widely adopted this isn't so much of 
a consideration, but was handy in the days that there were packages that 
could handle multiple layers in an EXR but didn't understand the 
multiview extensions.

From an efficiency point of view, EXR stores channel data in 
alphabetical order by channel name. Since stereo views of the same pass 
are often almost identical, storing them adjacently may well help 
compression and give a smaller file. However, we didn't really extensive 
testing of this with typical imagery.

It would be worth looking at the EXR-2.0 code base (currently the 
develop branch on github) for stereo images too: that doesn't store the 
view in the channel name, instead storing it in the part's 'view' 
attribute. ImfPartHelper.h generates correct multiview channel names and 
recommended part assignments for both (EXR-2.0) and single part 
(EXR-1.7.X compatible) multiview EXRs, so you should be able to handle 
both versions with very little extra code.

Peter


On 04/09/2013 12:29 AM, Ton Roosendaal wrote:
> Hi all,
>
> There's a project in Blender to get (at last ;) good stereo support integrated 
> for the render/composite/display pipeline.
>
> Obviously we'll try to align well to the proposed OpenEXR spec of it.
> http://www.openexr.com/MultiViewOpenEXR.pdf
>
> My question is why the authors proposed to insert the view 'layers' on the 
> lowest level in the structure. Was this for compatibility, or other reasons?
>
> In Blender internally, we use a list of "render layers" with a list of "passes" 
> which each have the channel buffers. The paper proposes to then add views as follows:
>
>     render_result -> layers -> passes -> views -> channels
>
> An other way would be to move views to the top level:
>
>     render_result -> views -> layers -> passes -> channels
>
> It probably doesn't matter much (exr file read can map to how we store it 
> internally), but I'd be very interested to hear if there's people here who did 
> a stereo migration for their pipelines and can share some experience.
>
> Thanks,
>
> -Ton-
>
> ------------------------------------------------------------------------
> Ton Roosendaal  Blender Foundation   address@hidden    www.blender.org
> Blender Institute   Entrepotdok 57A  1018AD Amsterdam   The Netherlands
>
>
> _______________________________________________
> Openexr-devel mailing list
> address@hidden
> https://lists.nongnu.org/mailman/listinfo/openexr-devel