[XviD-devel] Qpel bug - please help...

[Michael Militzer]

Hi,

----- Original Message -----
From: "Radek Czyz" <radoslaw@syskin.cjb.net>
To: <xvid-devel@xvid.org>
Sent: Sunday, December 08, 2002 1:42 PM
Subject: [XviD-devel] Qpel bug - please help...


> Hey everyone.
>
> I've been tracking qpel artifacts ever since qpel appeared - but with
> no success.
> Let me explain: qpel gets decoded wrong by different decoders. It's
> difficult to tell how wrong, but it's mostly visible when new
> keyframe changes the picture.
>
> I've hacked motion estimation to create a clip when the problem is
> clearly visible.
> Please load a short clip from http://noa.sm.pl/~syskin/test.avi .
> Second row of macroblocks always has the vectors of (2, 0) (in qpel
> resolution). Second colum of macroblocks always has (0,2). Diagonal
> macroblocks have (2,2).

nice clip, this is helpful...

> This is what happens: if xvid decodes the clip, it's ok.
> If ffdshow decodes the clip, we can see that all 3 types of
> interpolation are decoded completely wrong, which causes artifacts
> over time (note: xvid iDCT, but it's the same with other iDCTs, so
> it's not an iDCT problem).
> Now the funny thing: if DivX5 (version 5.01, I don't have any other)
> decodes the clip, only HV interpolation (diagonal macroblocks) are
> wrong - H and V are correct. (?)

ok, I just checked for myself and have the same impressions. XVID decodes
correctly, DivX only shows a problem in the HV filtered part and ffdshow
always (H,V,HV) introduces artifacts.

I also quickly checked with ms fdam and the clip is decoded with similar
artifacts than what ffdshow does. Obviously we (and DivX btw) have a
compliance problem. Since the problem is not clearly visible (before
decoding 100 frames), I suppose it could be a rounding (?) problem int the
low-pass filtering code. I'll dig into this when I'll have some time
again...

> I'd really appriciate if someone could test decoding with more
> decoders - reference ones would be great. I really don't know if it's
> xvid's problem or ffdshow's and divx5's.... But it's a problem, I'm
> sure of it.

yes, it really is a problem. Hopefully the current bug introduces just a bit
of smearing with compliant decoders and no severe artifacts, but anyway:
everyone (should) know that the dev-tree is unstable and therefore no good
choice for archiving purposes.

bye,
Michael