[XviD-devel] Quarterpel problem.

[Christoph Lampert]

Hi,

On Tue, 8 Oct 2002 jeansonh@admtek.com.tw wrote:
> [...]
> I also modified the code to decode B-VOP with quarterpel. If combined the
> core
> with xvid_decraw frontend, the decoding process 'SEEMS' okay, but failed
> if compiled as directshow filter... :( The source sequences are generated
> from
> DivX 5 CODEC...
> 
> And... in standard, B-VOP with quarterpel in forward, backward, and
> bi-directional
> mode should be interpolated within the 16x16 pixels. But in DivX
> implementation
> it looks like using 4 8x8 pixel blocks to do the interpolation. Anyone
> knows
> whether it's correct or that's my misunderstanding?

Where does it say 16x16? I just found 

7.6.9.5.3 Generation of prediction blocks 
Motion compensation for luminance is performed individually on 8x8 blocks
to generate a macroblock. [...] Note that in the case of quarter sample
mode, the motion compensation of four 8x8 blocks, using the same motion
vector for each block, does not lead to the same result as the motion
compensation of a single 16x16 block using the same vector. This is due to
the block boundary mirroring described in subclause 7.6.2.2.

> How could I provide my codes? :)

Easiest method at the moment is to send a patch to xvid-devel.


gruel