[FFmpeg-devel] pixel format convertion yuv422p->yuv420p->yuv411p->yuv422p degrade picture

[Siarhei Siamashka]

On Tuesday 15 July 2008, Maksym Veremeyenko wrote:
> Michael Niedermayer ???????(??):
> > On Tue, Jul 15, 2008 at 04:22:33PM +0100, M?ns Rullg?rd wrote:
>
> [...]
>
> >> Without thinking a little harder, I don't know whether such a conversion
> >> sequence is possible without such drift.  It will always be lossy, as
> >> I'm sure you already realise.
>
> I did cyclic test for:
>
>   +->422->420->422->411->422-->+
>
>   +----------------------------+
>
> no visial degrades found.
>
> > I suspect that dithering could reduce the drift, for normal interpolation
> > based filters but running 50 converters in a row is a stupid thing to do.
>
> Such condition possible in:
> Uncompressed (4:2:2) -> DV25 (4:2:0) -> DVCpro25 (4:1:1) -> Uncompr
>
> 50 iterations was only for proving that convertions degrade picture...
>
> So, as conclusion, that a normal?

BTW, I just tried your script on "uniform white", "gray gradient" and "some
random picture from my collection" images. The problem is not reproducible for
me (the difference between first and last images from the sequence is not
visible at all). Is there anything special in your setup? Anyway, now I doubt
that it is the effect of rounding. More likely something may be screwed up
and you get uninitialized zero bytes in the chroma data. Can you try building
FFmpeg without any assembly optimizations and rerun the test?

-- 
Best regards,
Siarhei Siamashka