[XviD-devel] inter/intra screener

[Andrew Voznytsa]

IMO it does not depend if (A)SP uses spatial prediction in Intra case or 
not. SAD / SBE shows you degree of temporal correlation. In case of low 
temporal correlation it's good place to check Intra modes.

I propose the following algorithm prototype:

1) If best Inter cbp is 0 - skip Intra checks at all.
2) compare SBE and best Inter SAD. For example (if up and left edges 
available): SAD / 256 * 0.5 > SBE / 32  then try to check Intra modes.
3) test only one Intra mode (i.e. no AC/DC) and compare (RD-) with best 
Inter mode. If numbers are more or less close - do Intra refinement 
(test other Intra modes).

Best regards,
Andrew Voznytsa



Radek Czyz wrote:

> Hi
>
> Andrew Voznytsa wrote:
>
>> you may try to use MB boundary error. See 
>> ftp://jvttest:tester@standards.polycom.com/2003_12_Waikoloa/JVT-J033r1.doc, 
>> section 3,*/ "Selective intra mode decision".
>> IMO this should work for MPEG-4 ASP too.
>> /*
>>
>> Best regards,
>> Andrew Voznytsa
>
>
> Hehe yes, they're doing exactly the same thing I'm doing and for 
> exactly the same reason. The only problem is that in mpeg-4 (a)sp, 
> intra is almost not predicted from other macroblocks - so the boundary 
> error has no effect on intra's rate.
>
> They do however confirm that I should try to estimate the rate of 
> intra and compare it to the (known) rate of inter. Right, I knew that 
> ;p and I'll try it soon-ish.
>
> As for now, I can help with early-stops - the minimum number of bits 
> taken bu INTRA is 8 (minimum cbp 2, minimum mode 5, and one ac/dc 
> prediction bit) and I can add this eight to our known "so far" as soon 
> as possible.
>
> Radek
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel