[XviD-devel] Cartoon mode

[Michael Militzer]

Hi,

Quoting Christoph Lampert <chl at math.uni-bonn.de>:

> On Thu, 17 Apr 2003, Radek Czyz wrote:
> 
> > > I'm
> > > still wondering why Qpel is sometimes worse and often better than
> halfpel. I
> > > have an idea and I'm currently investigating this.
> > 
> > I came up with an answer recently, when I was playing with my LOTR
> > DVD. It turns out that qpel can help - really help - when the image is
> > sharp, while it doesn't help when the image is blurry.

hm, I'm not so sure about sharpness: I just did a test and encoded one of my 
test clips (which are qpel friendly) but applied a "blur more" in VirtualDub as 
prefilter. Result: Qpel still performed much better than halfpel, comparable to 
the difference without the bluring filter. But further investigations are 
needed...

> Which raises new questions: Is this because Qpel ME is bad for smooth
> images, or because QPel's Halfpel interpolation doesn't smooth the image
> as much as Halfpel's? Or really a Qpel-Problem? 
> 
> Or maybe neither, and Qpel _is_ good, but Halfpel is good already
> (good "enough"), so Qpel's longer vectors kick in? But then of course,
> only filesize should increase, PSNR shouldn't change.

it's not a "longer vector" problem. In high quality encodings (quant < 5) that 
we are usually doing, the bits to code the vectors do not play a major role. 
The dominant part of the overall bit rate is made of the quantized 
coefficients. And in my tests I've discovered that Qpel sometimes leads to an 
increase in texture bits (in comparison to halfpel) which would explain why 
Qpel sometimes performs significantly worse than halfpel. However this 
obervation still does not solve the question: why does qpel often reduce 
texture bits but sometimes leads to an increase?

bye,
Michael