[MPlayer-users] RFC: docs update for "how to create a high quality DVD rip"

[Jason Tackaberry]

On Mon, 2004-06-07 at 15:46 -0500, Wayde Milas wrote:
> Raid 5, linux kenel raid. So 7x250 useable 1 hot spare, minus formatted
> loss. I use Reiserfs on it.

Very nice.  I wish I had the money for that much space. :)

> I'm doing quality tests on a 19 inch flat panel lcd. I dont have a
> monster hdtv to test display on. I'm wondering how it looks blown up
> huge.

I notice about the same amount of artifacting when my face is 1 foot
away from my 15" LCD laptop screen as I do when I watch the video on my
52" rear projection about 9 feet away.  Perhaps slightly more on the tv,
but it's close.

> Yes I read this. Its much better than the default on btw, and I added
> some comments at the end of my last email.

Ok, thanks.  I'll do a bunch of testing with *cmp=(2|3) so I can
convince myself it's superior. :)

> does that noise bit add in the 2:1:2 that was taken out?

It adds a lot more, in fact.  hqdn3d=2:1:2 is very little.  Too little,
according to Rich, but I'd rather be overly conservative.  With -vf
noise=8ah:5ah (or so), it has about as much noise as the DVD does, and
it masks certain artifacts.

It all comes down to perception.  Obviously by adding random noise
you're lowering the PSNR, but it _looks_ a lot more convincing.

> wondering if you should depend on spp, or just up the bit rate? Ie,
> should I be doing my tests by running the stream through spp for
> viewing, or just viewing it "as is"?

Upping the bitrate stops helping when every frame is already using qp=2.
spp helps clean up the image to a certain extent, and adding a bit of
noise helps even more, IMO.

> Of course. I'm just afraid of it taking out detail that it thinks is
> noise.

If it does, it isn't likely to be noticeable except for larger denoise
parameters.

> feel for what is actually going on. Finding the right bit rate seems to
> be more of a art than a science.

I agree, it's more an art, because of this weird word called "quality"
that seems to be immeasurable.  Or at least highly debatable. :)

> What I was considering was taking a few chunks of a mpeg2 stream,
> running basic q=2 high quality fast encoding on the chunks and recording
> the psnr values in a script. then going BACK over those chunks in a
> script with full options and altering the bitrate till the psnr values
> are within a certain percentage.. maybe 5%.

I think relying on PSNR to find your optimal bitrate is a mistake.  At
higher bitrates, PSNR appears to be much less meaningful.

Write a script that encodes your clips at various bitrates and view
them.  Have your script also take several still frames at various
positions so that you can do an A/B toggle.  (Doing A/B on stills is
less telling, but at least it's something.)

In fact, I've already written such a script (brutal hack) to do some
lavc/xvid comparisons.  It's definitely interesting to view the results.

> Er what did he do last night? whats cmp=10?

Read the last few emails in the "lavc vs. xvid" thread.  I uploaded a
clip where lavc showed quite a lot of blocky artifacts, and he added a
new cmp func called NSSE that worked wonders.

> Is that the 10 function? In the past higher psnr values have always
> looked better, or the same to me. I can't remember when a lower one
> actually looked better...

PSNR might be the best quantifiable way to measure quality, but quality
is also a matter of perception.  Adding noise reduces PSNR but masks
those pesky blocks.  The brain has an easier time ignoring random noise
than dancing blocks.  (At least my brain does.)


> I think some of the asm optimizations for p4 are fubared somewhere :P

I think most of the asm optimizations are MMX, with work the same on
both AMD and Intel.  I don't know enough about mplayer to explain the
difference.

Cheers,
Jason.