[XviD-devel] DivX5.03 Nth Pass encoding

[Christoph Lampert]

On Mon, 27 Jan 2003, Cami wrote:

> Looking through the divx 5.03 changelog:
> 
> ::::
> .. Nth PassT encoding. Another new addition to the rate control arsenal, Nth PassT encoding takes advantage of information
> available from previous analysis (or analyses) of the video sequence. Nth Pass encoding uses information garnered from previous
> passes through the video sequence, which can improve quality with each pass. There are four main steps to Nth Pass encoding:
>   1.. Analysis (the first pass)
>   2.. Map generation (where the results of analyses are processed)
>   3.. Strategy generation (strategist examines map of the video sequence and creates a strategy that will best achieve the three VBV
> RC requirements)
>   4.. In-loop nth pass rate control (uses information from the map and strategy to effect rate control that best fulfills the three
> requirements)
> ::::
> 
> Can anyone tell me if xvid having such a feature will
> improve image quality? or is it not nessasary / wont make any difference?

I don't see the difference to two-pass, except for fancy "strategy"-talk,
which doesn't mean a thing.

Possibly three-pass would be better than two, if we analysed e.g. blocking 
artefacts from second pass. But if it's worth the extra time? I don't
know.

What might be cool is a new feature from transcode: "On the fly parameters
changes". Create a logfile with encoding setting in advance (not only
"start of credit, end of credit", more like a movie script, including
not only quantizers, but also things like motion search method, GMC,
maybe quant matrix (ratecontrol could of course still modify
quantizers).
In a faster "0th-pass" we could analyse the material for this, also
detect keyframes etc. In particular, since we don't encode, we could use
multiple-frames read-ahead, etc. do decide on b-frames. Then we do
ordinary two-pass with this "script".

But I don't think that's what DivX does...

gruel