I'm trying to make some very high quality MPEG-2 coded videos
of material that is somewhat different than a "natural image"
and I've exhausted all regular means of improving the image
quality (best encoder/highest bitrate/etc.),

I'm trying to figure out if I should attempt to optimize the quantizer
matrix. From the Y signal standpoint, I have a lot of bright points
on the order of 3 to 12 samples wide set against a dark background
(the video is of fireworks). There is not a whole lot of subtle
gradient
in the background that need to be retained. In fact, most of it is
plain black. I mainly need to render the high contrast areas as
efficiently as possible and with Gibbs ringing that is tuned to be
the least obnoxious given the subject matter/HVS.

Is this worth trying to customize my quantizer for the content? If
so, and I have a great quantizer matrix for regular content, is
there an analytic procedure to bias it in favor of a particular
frequency response or is it all guessing bands of numbers and
grading the result aesthetically?

It would be great if there was a utility that made sophisticated
assumptions about the HVS and could process a video to come
up with an optimal quantizer matrix.

Potatophysics wrote:
> Is this worth trying to customize my quantizer for the content? If
> so, and I have a great quantizer matrix for regular content, is
> there an analytic procedure to bias it in favor of a particular
> frequency response or is it all guessing bands of numbers and
> grading the result aesthetically?

Yes, but without knowing what encoder you use it's difficult to advise
further.

CinemaCraft Basic is a wonderful encoder with exactly these sets of
controls, and you can dial the bias in manually if you're not happy
with the presets.  It's under $60.

Jim Leonard wrote:
> Potatophysics wrote:
> > Is this worth trying to customize my quantizer for the content? If
> > so, and I have a great quantizer matrix for regular content, is
> > there an analytic procedure to bias it in favor of a particular
> > frequency response or is it all guessing bands of numbers and
> > grading the result aesthetically?
>
> Yes, but without knowing what encoder you use it's difficult to advise
> further.
>
> CinemaCraft Basic is a wonderful encoder with exactly these sets of
> controls, and you can dial the bias in manually if you're not happy
> with the presets.  It's under $60.

I'm using Cinemacraft Basic. I also have the current TMPG encoder
product
which allows for completely custom quantization matricies. I tested
Cinemacraft
with a large number of values for the Flat part priority and I acutally
haven't seen
that much of a difference in the final result so I've been looking at
the
quantizer matrix. It seems that only Cinemacraft Basic doesn't allow
for fully a
fully custom matrix and that one has to go to SP to get that feature.

"Potatophysics" <moored@secac.com> wrote in message 
news:1169150707.973434.59580@q2g2000cwa.googlegroups.com...
> I'm trying to make some very high quality MPEG-2 coded videos
> of material that is somewhat different than a "natural image"
> and I've exhausted all regular means of improving the image
> quality (best encoder/highest bitrate/etc.),
>
> I'm trying to figure out if I should attempt to optimize the quantizer
> matrix. From the Y signal standpoint, I have a lot of bright points
> on the order of 3 to 12 samples wide set against a dark background
> (the video is of fireworks). There is not a whole lot of subtle
> gradient
> in the background that need to be retained. In fact, most of it is
> plain black. I mainly need to render the high contrast areas as
> efficiently as possible and with Gibbs ringing that is tuned to be
> the least obnoxious given the subject matter/HVS.
>
> Is this worth trying to customize my quantizer for the content? If
> so, and I have a great quantizer matrix for regular content, is
> there an analytic procedure to bias it in favor of a particular
> frequency response or is it all guessing bands of numbers and
> grading the result aesthetically?
>
> It would be great if there was a utility that made sophisticated
> assumptions about the HVS and could process a video to come
> up with an optimal quantizer matrix.
>

well, with video it is harder, but for images the algo is not that 
difficult.


actually, imagining a good set of values for a quantizer matrix shouldn't be 
too difficult anyways.

note that the right-hand-side is for higher frequencies (horizontally), and 
the bottom is for higher frequencies (vertically).

one then just imagines the energy distributions, and tries to come up with 
good numbers.


I would guess:
well, luckily you don't need that much for most of the image, so fairly high 
values can be put in much of the middle of the matrix (lower-left corner to 
upper right), the values in the opper-left should be fairly low for general 
appearance (but may be relatively higher than the lower-right corner).

probably for the lower right, values should be fairly low.

you would probably want to avoid any harsh and sudden jumps or drops in the 
numbers though, so trying to maintain some kind of curve-gradiant may make 
the most sense.


using this, you can probably adjust things so that it looks about how you 
want.

Potatophysics wrote:
> > CinemaCraft Basic is a wonderful encoder with exactly these sets of
> > controls, and you can dial the bias in manually if you're not happy
> > with the presets.  It's under $60.
>
> I tested
> Cinemacraft
> with a large number of values for the Flat part priority and I acutally
> haven't seen
> that much of a difference in the final result so I've been looking at
> the

You don't use larger values for sharpness, you use smaller values.
Larger values bias the flat parts; smaller values bias the sharp parts
(but may result in banding in flat parts).  If you want all sharpness
at the expense of the flat parts, you choose values of 16 or lower for
"Quantization Characteristics: Flat Part Priority".  Also try a few
experiments with the presets for Animation if that's what you're trying
to compress.

> quantizer matrix. It seems that only Cinemacraft Basic doesn't allow
> for fully a
> fully custom matrix and that one has to go to SP to get that feature.

Correct.