[FFmpeg-devel] H.264 + PAFF: BBC HD recording shows extreme interlacing artefacts

Reinhard Nissl rnissl
Sun Oct 28 10:07:32 CET 2007


Hi,

Michael Niedermayer schrieb:

>> Strange is that splitting the image into images which just contain odd
>> or even lines respectively doesn't give you two "crisp" images. After
>> splitting the odd lines image for example once again, I've got a crisp
>> image and one which looks like luma and chroma don't match.
>>
>> This leads me to the conclusion that there might still be a decoding
>> issue with H.264 + PAFF.
>>
>> Shall I upload this recording for further investigation?
> 
> no, there is no problem with decoding, the problem is in the yv12->rgb
> convertion, the fields have to be converted seperately not like a frame

Just to clarify, this sounds like taking the snapshot as PNG is buggy.
But the same artefacts are also visible when using libxv (by xine's
option -V xv) which does the colorspace conversion in hardware.

Is it that libxv must be informed about how to do the colorspace
conversion or must hardware colorspace conversion be turned off in such
a case respectively?

As the image looks still wrong when using -V xshm, which should do the
colorspace conversion in software, it seems that this part of code needs
some work, too.

> you can eiher send us a CLEAN patch which changes ffmpeg.c / ffplay.c
> to support that or better implement it in libavfilter
> 
> to see some example on how to do it see libmpcodec/vf_scale.c in mplayer
> svn

MPlayer dev-SVN-r24824-4.2.1 (C) 2000-2007 MPlayer Team
[snip]
VDec: vo config request - 1440 x 1080 (preferred colorspace: Planar YV12)
VDec: using Planar YV12 as output csp (no 0)
Movie-Aspect is 1.82:1 - prescaling to correct movie aspect.
VO: [xv] 1440x1080 => 1964x1080 Planar YV12

Mplayer shows the same issue, at least when using xv. I assume that the
above suggested changes are only meant for getting a PNG image, isn't it?

Bye.
-- 
Dipl.-Inform. (FH) Reinhard Nissl
mailto:rnissl at gmx.de




More information about the ffmpeg-devel mailing list