[XviD-devel] Re: XviD-devel Digest, Vol 14, Issue 5

[Christoph Lampert]

On Tue, 9 Mar 2004, JerryChen wrote:

>Hi all:
>    I have joined the devel maillist for two weeks and found that
>this maillist is not as active as i thought. Because of the lack of
>documents, it seems to be difficult to do something helpful for
>new developers.
>I wish somebody could provide some useful document for the newbies.

>_ [...]

> 
>     I'm a researcher on computer vision and interated in the implementation 
> of motion estimation. Maybe i should spend more time to look into the 
> source code.

Hi,

the list is very quiet, that's true. One reason might be that XviD is very 
close to its 1.0 release, and new features or big changes are postponed 
until after. 

Motion Estimation is an adapted PMVfast/EPZS zonal algorithm. There are 
some papers by Alexis Tourapis available online to explain the general 
method. In 2 words, it's just a gradient decent in image space, 
with the starting point chosen from a set of predictors that depend on 
neighbour blocks. 
The functional optimized is   SAD(v) + lambda(q)*BITS(v)   where 

SAD(v) = Sum of Absolute Differences between current block and block in 
         reference image with displacement v

BITS(v) = bits needed to encode the motion vector v in the bitstream

lambda(q) = a parameter derived in the quantization stepsize q. 

The background for this are papers on Rate-Distorsion optmization e.g. 
from Thomas Wiegand: 

http://bs.hhi.de/~wiegand/ICG-Project-RDO.html

gruel