ITU-T SG16 Q.6 VCEG       sept. '01      Document VCEG-N74.txt


From: Thomas Wedi [wedi@tnt.uni-hannover.de]
Sent: Wednesday, September 26, 2001 1:01 AM
To: Gary Sullivan; Wiegand,Thomas
Cc: Wien, Mathias
Subject: Re: Concerning VCEG-N51

Gary,

if the whole image is sampled up to the final resolution, only the
intermediate temporary frames in the interpolation process need longer
wordlength. 

This means:

for 1/4-pel in the encoder:
===========================
- in the current implementation the 1/4-pel resolution is stored in the
encoder with the same wordlength than before
- only the intermediate 1/2-pel resolution during the interpolation
process is stored temporarily with a higher precision

for 1/8-pel in the encoder:
===========================
- in the current implementation only the 1/4-pel resolution is stored
with the old wordlength
- only the intermediate 1/2-pel resolution during the interpolation
process is stored temporarily with a higher precision
- the 1/8-pel values are created on the fly by bilinear interpolation if
they are needed

for 1/4-pel and 1/8-pel in the decoder:
=======================================
- only the original resolution is stored
- all subpixel values are generated on the fly by the direct
interpolation.



You are right, this is an complexity impact in the encoder. But from my
point of view it is not significant, because the frames with longer
wordlength are stored temporarily during the interpolation process. On
the other hand you have a nice complexity reduction in the decoder,
because you only store the frames at their original size in the decoder.
Furthermore the number of multiplications in the decoder is decreased by
using the direct interpolation (VECEG-M46).

Peter List brought up an other issue. In his implementation he wants to
store only the 1/2-pel frames in the encoder. He performs the 1/4-pel
interpolations on the fly. This will be a problem with the current
solution, because then he has to store the frame with the longer
wordlength.

If this is not acceptable, we should think about an alternative solution
in order to resolve this problem by still having the advantage of the
direct interpolation at the decoder. One approach in this direction is
proposed in VCEG-N31. 

I am planning to bring more results and new approaches concerning this
point to the next meetings.

If it is reasonable, please feel free to present my comments to the
group.

Best Regards,

Thomas


Gary Sullivan wrote:
> 
> This increases wordlength in the encoder.  Doesn't that
> have a signficant encoder complexity impact?
> 
> -Gary
> 
> +> -----Original Message-----
> +> From: Thomas Wedi [mailto:wedi@tnt.uni-hannover.de]
> +> Sent: Monday, September 24, 2001 12:26 AM
> +> To: Gary Sullivan
> +> Subject: Re: Concerning VCEG-N51
> +>
> +>
> +> All pixels are rounded only after the last interpolation step.
> +>
> +> Thomas
> +>
> +> Gary Sullivan wrote:
> +> >
> +> > +> Thus, I changed the rounding procedure of the subsequent
> +> > +> interpolation in a way, that only one rounding step at
> +> > +> the end of this subsequent scheme is performed.
> +> >
> +> > How?
> +> >
> +> > -Gary
> +> >
> +> > +> -----Original Message-----
> +> > +> From: Thomas Wedi [mailto:wedi@tnt.uni-hannover.de]
> +> > +> Sent: Friday, September 21, 2001 6:28 AM
> +> > +> To: Gary Sullivan; Wiegand,Thomas
> +> > +> Subject: RE: Concerning VCEG-N51
> +> > +>
> +> > +>
> +> > +> Gary,
> +> > +>
> +> > +> If you have a direct interpolation scheme, you have only one
> +> > +> rounding step at
> +> > +> the end of the interpolation. If a subsequent interpolation
> +> > +> scheme is used, it
> +> > +> is possible to have a rounding step after each interpolation
> +> > +> step. E.g. for
> +> > +> 1/2-pel: one rounding after 1/2-pel interpolation and one
> +> > +> after 1/2-pel
> +> > +> interpolation. But then you have a mismatch between these
> +> > +> interpolation
> +> > +> schemes, because one scheme has only one rounding step and
> +> > +> the other one has
> +> > +> more rounding steps. Thus, I changed the rounding procedure
> +> > +> of the subsequent
> +> > +> interpolation in a way, that only one rounding step at the
> +> > +> end of this
> +> > +> subsequent scheme is performed. Then you have no mismatch.
> +> > +>
> +> > +> Thomas
> +> > +>
> +> > +> Gary Sullivan <garysull@microsoft.com> schrieb:
> +> > +>
> +> > +> >
> +> > +> > Thomas,
> +> > +> >
> +> > +> > One question: What is it that you refer to as the solution
> +> > +> > to the rounding problem in direct interpolation?
> +> > +> >
> +> > +> > -Gary
> +> > +> >