[XviD-devel] XVID 1.0 + c't

[Michael Militzer]

Hi,

Quoting Edouard Gomez <ed.gomez at free.fr>:

[...]

> Btw I'm starting playing with p-domain rate control, at this moment i'm
> just checking if the Zhihai He paper is right. I have doubts on his
> formal/theoretical foundations but i think the method is right in
> practice.

Ed, I already have a p-domain rate control implementation ready and it
works quite well. However there are fundamental changes needed in xvid
structure so I didn't plan to integrate the code before the new api
settles down a bit. Also p-domain rate control is noticeably slower than
our current rate control and as Christoph mentioned, we don't really make
use of the extra accuracy from p-domain rate control (since xvid right
now mainly targets storage applications, small rate control buffer sizes
are not that crucial than in a live streaming environment).

> I have some points to clarify tho:
>  - does the model need to be per frame type ? (the three parameters we
>  are supposed to update while encoding)

I'm not sure which algorithm you refer to exactly (Mr. He has published
quite a few papers) but in general the rate curve characteristics are
quite different for different frame types.

BTW: I've looked at the first p-domain patch you sent to the list and I
wonder why you count the zeros in mbcoding.c? You need the information
about zeros _before_ coefficient coding (and also before the quantization
step) otherwise it's pretty useless for a single-pass rate controller.
 
>  - Though the accuracy increases, what will be the use of this accuracy
>  boost ? better strict CBR... former CBR algorithm was lean and very
>  simple, and quite accurate. Better control in 2pass?... you see that if
>  this p-domain is useful, we have still to find out how to use it right.

Our current rate control algorithm is indeed lean and simple. However I
wouldn't say it's accurate. As we only control frame quantizer, actual and
target frame sizes are often not very close. Of course overall performance
is quite good but this is because we use a relatively large rate control
buffer.

The higher accuracy of p-domain rate control is especially helpful in real-
time low-delay video transmission, but does not give a major improvement
for video storage (at least not as long as you don't exploit the rate
control buffer for R-D optimized frame bit allocation - but this is not
that easy).

bye,
Michael