[XviD-devel] HUGE ME regressions between beta1 and beta2

[Radek Czyz]

Hi,

Michael Militzer wrote:
> That should be easy to find out: if there's no regression with VHQ>0, then
> the degradation on some sequences likely comes from mode decision. That
> should be easy to fix then by applying additional weights after vector
> search. Also, it may explain why the problem slipped through unnoticed - 
> most people use VHQ I guess.

Yes, I just confirmed that at defaults (with VHQ1), the quality is a bit 
higher accross the board. It doesn't exactly look like it because 
filesize is bigger, but apparently PSNR offsets it. I get the same 
graphs you got.

I will now try to tweak all constants for best PSNR with VHQ0. I will 
also try to find why there's such a big slowdown - Skal's theory might 
be correct, and in that case we're doomed - we can't get this 5..10% 
speed back. I have a hope that explaination is different and some 
early-skip thresholds no longer work correctly, with different levels of 
total SAD. If so, we can return to old speed (hopefully still keeping 
new psnr).

Radek