On Mon, 2 Jun 2003, Steven M. Schultz wrote:
>> From: Daniel Pittman <[EMAIL PROTECTED]>
>>
>> I have been using mpeg2enc from the current CVS tree together with
>> transcode to encode DVD compliant MPEG-2 data from PAL DV input.
[...]
>> The one big problem that I do have with the toolset is that I get
>> ghosting in very dark scenes, to the point of making some stuff
>> *very* uncomfortable to watch -- like seeing it through a heavy, dark
>> fog or something.
>
> Hmmm, the common problem mentioned has been "splotches of grey" in low
> light scenes - hadn't heard 'ghosting' mentioned before.
That's the same problem, just differently described. If you look at each
individual P or B frame, there are "splotches of grey" where, er, things
were in frames from the previous I frame.
Each time something moves it drops a new grey splotch and the previous
splotch fades a bit, so you get an odd fading trail, or a grey blur,
from a moving object -- depending on how fast it moves.
>> The mpeg2enc command line[1][2] is:
>>
>> ] mpeg2enc -v 0 -f 8 -b <bitrate> -F 3 -n p -a 2 -o <output> \
>> -a 2 -q 2 -I 1 -r 24 -4 2 -2 1 -N 1.5 -Q 1.5
>>
>> The bitrate is usually around 5500 to 7000, and the ghosting doesn't
>> seem to be greatly affected by that, sadly.
>
> Hmmm, what about trying a little less agressive -N option? Perhaps 1.0
> or 0.8 would help (but not increasing the high frequency quantizers as
> much).
Just to make sure I have this right, a higher quantization means a
*lower* degree of information recorded for that macroblock/frame, right?
I can see how this might make a significant difference. I will hunt
through my current set of DV sources and see if I can find a sample that
has this issue. Sadly, I only watched 'Baby Cart in Hades' after I
deleted the raw DV; it shows this really, er, well.
> The other thing that can cause artifacting (at least it was mentioned
> as having that possibility) is the -Q option. I'd try leaving that out
> and seeing what happens.
*nod*
> -q 2 is extremely low - a value of 5 or perhaps 4 is about as low as
> I've ever gone. A "-q 4 -N 0.6" produces very good output and
> preserves more of the detail than -N 1.5 can.
OK, that makes sense. I figured that I would need to play with the
numbers to get -q 2 working sanely, but it actually seemed to be fine,
so I left it alone.
I have tried, previously, to find a rough guide to what -q value to give
for fitting a given length of time to a target size, but failed.
Is there any rough guideline you can suggest for trying to pick the
quantizer?
>> So, is there anything that can be done to improve this? This is the
>> one
>
> Is the effect reduced if the DNR is not performed? Sometimes the
> denoising filters have an effect that sounds like what you mention
> seeing.
Nope, that seems to be pretty much irrelevant. I will omit it from my
testing, though, when I grab a suitable sample segment.
> I'd try less aggressive hf "rolloff" (-N) and -q settings - could be
> that the ghosting is being caused by by an interaction there.
I will post a note on the list once I have done more testing so that
others can learn from this.
Daniel
--
I saw that most programmers never mature above the "see jack run" level.
My pals at the [suppressed!] Comp Sci Dept scoff at the estimates I make, but
I never underestimate -- they always do. I think of the big picture, the 75%
that remains after the code "works".
-- Erik Naggum
-------------------------------------------------------
This SF.net email is sponsored by: eBay
Get office equipment for less on eBay!
http://adfarm.mediaplex.com/ad/ck/711-11697-6916-5
_______________________________________________
Mjpeg-users mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/mjpeg-users
