Re[4]: [XviD-devel] New Motion Estimation (from sysKin) committed to branch

[Marc FD]

> > BTW, can you try a mode when Hinted ME would be used in one pass ??
> > i know it would be a bit slower but it would allow advanced
pre-processing.
>
> In fact, it's already there. P/B/I decision is based on a simple
> ME-like routine. It currently writes motion vector to MACROBLOCK array
> because it's needed for future get_pmv(), but doesn't write
> macroblock's mode (inter/intra) because it's not needed anywhere.

i'll look deeper in the source. it seems i'm closer and closer to the hot
spot
where Pre Processing should be done :)

>
> I'm not doing any simplier ME after this (but I could) because I don't
> expect to have this analisis done for second pass. No wait, maybe I
> really should do a special post-analisis search mode - deviation is
> already computed, as well as SAD of 0,0 vector and some other things.
> Will be faster, at least for firs pass/cbr/etc.

good

>
> As for pre-processing, I don't know what kind of preprocessing should
> be done when we have b-frames available - noise is updated every
> second/third frame anyway, while idea of loosing P-frame's details
> sounds horrible to me (B-frames won't re-create them because of higher
> quant).
> Well, I'm lame when it comes to stuff like this ;)
>

Pre Processing should be done on the INPUT material BEFORE any precise ME.
the use of a simple ME is to have very usefull informations, wich would
help to do a better filtering based on movement infos. something very
usefull
would be a value stored for each MB incremented each time the MB has 0,0 MV
and reseted each time it's INTRA coded.

segmentation could use these information to segment backgrounds for exemple.