[theora] Encoding 2-Pass VBR?

[tterribe]

> Timothy Terriberry estimates a simple self-training range-coding
> backend could be reasonably expected to decrease bit usage overhead by
> 15-20%.
> 
> Is it really true? I found one implementation....
> http://svn.xiph.org/trunk/ghost/libentcode/
> 
> Did anyone verified this gain? How much is the exact gain?

The number quoted there comes from the result of similar experiments
performed during the development of H.264 (they're in the JVT archives
somewhere, but it's been years since I actually looked for this particular
one). I think I originally told Monty "12-20%", but the exact number isn't
really important as it varies from sequence to sequence, and also depends
on how one sets up the context modeling, which is still more of an art than
a science. No one has measured an exact value for Theora, that I know of.