[XviD-devel] stereo video streams

[Christoph Lampert]

Hi,


sage weil wrote:

> Hi,
>
> I'm looking for a way to efficiently encode a pair of video streams 
> from a stereo camera.  A few obvious possibilities present themselves 
> but are not entirely ideal:
>
> - Put frames side by side and encode as usual.  Macroblock searches 
> won't find similar blocks from the other camera with a localized 
> search...

This will not make a significant difference to encoding both 
independently, you are right.

> - Interleave left and right frames.  If xvid searches more than one 
> frame back then this would work reasonably well, but I'm pretty sure 
> it doesn't? And if it did the motion prediction would end up way off, 
> resulting in suboptimal performance...

This might indeed work well with codecs that support more than 1 
reference frame (e.g. H.264 aka H.26L aka AVC aka MPEG4v10), but XviD is 
MPEG4v2 and there it's not possible. I would also strongly advise 
against hacking the format to support this, because it would need quite 
some new syntax elements and would be completely incompatible, causing 
major confusion.

> Does anybody have any thoughts about what might be necessary in terms 
> of code changes to make this work well?

My idea was to interleave left and right row by row and test encoding 
with or without interlaced DCT. You can still access them independently 
by using a twice as high stride value, and in areas where the 
displacement is low, it might be beneficial to
have both views as close to each other as possible.


gruel