[XviD-devel] lambda patch

[Michael Militzer]

Hi Radek,

indeed, it's a short patch ;) I have no objections against it: it's 
reasonable to give a higher penalty to 8x8 modes, especially at VHQ=0.
Also, I have no problem with switching back to a pure linear lambda
scaling also for the low quants. As said, it won't make much of a 
difference either way.

As a side note: I've also ran some tests recently because of the lambdas.
However, aside from 8x8/16x16 mode decision it also seemed to me that
inter/intra decision isn't working perfectly with VHQ=0. On some sequences
I had large gains from lowering the value of MV16_INTER_BIAS, e.g. to 400.
You may want to give it a try if it works for your test sequences too...

Regards,
Michael


Quoting Radek Czyz <radoslaw at syskin.cjb.net>:

> Hi everyone,
> 
> I finally found some time to tweak lambdas and mode decisions back.
> Effect:
> 
> - slowdown introduced by Michael's patch remains, but is now translated 
> to PSNR gain at all quants, so it's worth it.
> 
> - I changed lambdas to be linear for low quants too. I understand 
> Michael's reasoning for lower-than-psnr-best values (and probably agree 
> with the theory). However, I believe that such HVS experiments belong to 
> 1.2.-127 tree, and need some testing. With linear quants everywhere, at 
> least we can explain slowdown as PSNR-boosting.
> 
> - I tweaked constants for much better VHQ 0 mode decision.
> 
> Patch is short, I put it online at http://syskin.is.dreaming.org/lambda.txt
> 
> I strongly think this is the best thing to do for 1.1.0, and the best 
> starting point for 1.2.-127.
> 
> Comments? If no one writes anything (why that crossed my mind? lol) I'll 
> commit to both cvs head and 1.1.0. If you disagree, say so now.
> 
> Regards,
> Radek
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
>