OT: 4:4;4 THE FUTURE OF VIDEO??????

As someone who started in 3d Studio Dos, I have always been used to
uncompressed full RGB targa outputs-it wasn't until the animation
had to hit the video world that the Assembled sequence would have to
be compressed.

I am happy to see that new XDCAMs are now 4:4:4 (no chroma-re sampling)
This is a step up from 4:2:0, no doubt.

Anyhow, I digress, does anyone think that 4:4:4 will be the new standard
in the near future?

AFAIK, XDCAMs are MPEG2 based, and 4:2:2, so they do chroma sub-sample.

If you think about it, chroma sub-sampling is "just" old analogue compression, although now performed digitally. It works, but it's crude. We can do oh-so-much better by not decimating the chroma, and applying modern compression technology instead. I'm sure that will come. However, the other trend is to compress raw sensor data instead, which makes a lot of sense too.

Graeme

[www.nattress.com] - Plugins for FCP-X

The new Sony HDC F-950 looks interesting, this is 4:4:4


Graeme Nattress Wrote:
-------------------------------------------------------
> AFAIK, XDCAMs are MPEG2 based, and 4:2:2, so they
> do chroma sub-sample.
>
> If you think about it, chroma sub-sampling is
> "just" old analogue compression, although now
> performed digitally. It works, but it's crude. We
> can do oh-so-much better by not decimating the
> chroma, and applying modern compression technology
> instead. I'm sure that will come. However, the
> other trend is to compress raw sensor data
> instead, which makes a lot of sense too.
>
> Graeme

That's an F950 - getting on a bit now..... Not XDCAM.

Graeme

I must be tired..for two days I was all excited about a successor to the Sony F-330, F-350 camera.

I best learn how to read better!

Have you had snow yet in Ottawa?

Way colder than balmy Sarnia,
right across the river from Port Huron, Michigan!



Graeme Nattress Wrote:
-------------------------------------------------------
> That's an F950 - getting on a bit now..... Not
> XDCAM.
>
> Graeme

Yup, we had snow the other day! Lovely. I love autumn.

The thing with the traditional camera paradigm is that you use 3 chips, and the subsampling of the chroma was done after matrix and gamma and conversion to YCbCr because they didn't have the bandwidth to transmit without the chroma sub-sampling. Even with HD a single SDI link is 4:2:2 and two are needed for 4:4:4 and that's more expense. Of course, for the most part recorders have been limited to 4:2:0 or 4:2:2 (not to mention transmission) and hence little need for getting a higher level of quality out of the camera. The F950 doesn't have in-built recording, so it made sense to allow for the full signal out as recording technology advanced to allow for it.

With the more modern data-centric approach to cameras, and with advances in compression technology, it makes sense to just use better compression rather than put in a 4:4:4 to 4:2:0 or 4:2:2 conversion block in the camera, and with a move to raw recording like we do at RED with the R3D file, the whole concept of chroma sub-sampling becomes redundant.

Graeme

Just curious... The RED stores 12 bit linear bayer data from the camera's sensor, but wouldn't it be better to store the sensor data with a log curve rather than a linear curve?



www.strypesinpost.com

It all depends on how the compression works..... :-)

Graeme

A simplified explanation, perhaps?



www.strypesinpost.com

This is the link to the HD camera, it says it is HD

[www.expandore.com]




Graeme Nattress Wrote:
-------------------------------------------------------
> That's an F950 - getting on a bit now..... Not
> XDCAM.
>
> Graeme

Yes it's HD, but not XDCAM - no in-built recording. And it's rather old being used on the last of the new Star Wars movies.

Graeme

Yep, not XDCAM-but still 4:4:4;
which is much closer to what you get
when you click Render Scene from 3ds Max or Autodesk Maya.

(and I always love to compare apples to oranges) >:


Graeme Nattress Wrote:
-------------------------------------------------------
> Yes it's HD, but not XDCAM - no in-built
> recording. And it's rather old being used on the
> last of the new Star Wars movies.
>
> Graeme

The difference in a camera is that you have a prism that produces optical effects, and some 4:4:4 camera heads I've measured (HDC1500 ? comes to mind) have alignment issues with the three channels that mean although the output is 4:4:4, you couldn't in all honesty call the signal that quality.

Graeme

What I meant was, is it better to sample RAW bayer data at 12 bit log than 12 bit linear? Or is it because the A/D converter converts to 12 bit linear?



www.strypesinpost.com

AD's are indeed linear. It's also necessary to have the data in linear form for matrix based colour correction.

Graeme

Chroma subsampling, though analog color TV did similar, is a very smart idea based on human visual science. There are 3 cone types in the retina (peaking at 450nm, 530nm, 560nm, not exactly for Blue, Green, Red) and yet the optic nerve does not carry these 3 signals to the brain. It carries instead sum and difference signals. They make a Red vs. Green channel, a Blue vs. Yellow channel, and a Luminance channel. Furthermore nature devotes fewer nerves (implying less spatial resolution) to the first two color-difference channels than to the Luminance channel. Our eye does chroma subsampling, so why shouldn't our cameras?

Uncompressed 4:4:4 requires 50% more data than uncompressed 4:2:2. If all work and release will be uncompressed, then 4:2:2 is smarter than 4:4:4, because you can then trade the data difference to get 50% more pixels, or get 12-bit instead of 8-bit. But if work and release will be compressed, then Graeme is right that better compression can be calculated from the full 4:4:4 data than from the reduced 4:2:2 data, even if the compression aims at exploiting the eye's chroma subsampling.

Bayer's original claim that his GRGB patterned sensor replicated the eye is false. (His Kodak fantasy of the eye still kicks around the video community.) A camera's Green sensing must not be confused with its Luminance sensing, because a camera's Red sensing also figures significantly in this. Nor are there Red or Blue signals leaving the eye. Red vs. Green ain't Red. Blue vs. Yellow ain't Blue.

Graeme is right that analog video made a mistake of constructing its chroma and luminance signals after gamma correction, but I don't agree with his answer to strype's last question. Even 10-bit log luminance carries more visual information (more discriminable levels) than even 14-bit linear luminance. If we could get 10-bit log R, G, B from the camera, then for at least some color calculations, such as for luminance on a grey scale, we get better answers by converting the log back to linear, calculating, and then converting back to log than we would from calculating directly from even 14-bit linear R, G, B from the camera. It requires more analysis whether this advantage holds up for many more color calculations.

Dennis Couzin
Berlin, Germany

>Our eye does chroma subsampling, so why shouldn't our cameras?

Image post processing. Any VFX guy will say that a heavily chroma subsampled image produces jaggies on the key. You never see jagged chroma subsampling artifacts in the real world. Basically if you plan to do VFX, you are shooting yourself in the foot if you shoot to a chroma subsampled image. For news casting or quick turnaround events, where storage and fast delivery is your main concern, subsampling may not be a bad thing.

My original question was why linear and not log. But since the data came from a 12 bit AD converter as lin, you can't get better than that without changing the hardware.



www.strypesinpost.com

strypes, very true that chroma-keying requires that the chroma resolution match the luminance resolution. I should have qualified my answer to non-VFX works. Isn't this the majority of video?

Concerning the log vs. linear question, I used 10-bit log vs. 14-bit linear just as illustration. Similarly, even 8-bit log gives a visually smoother luminance scale than even 12-bit linear. In the words of color scientist Michael Brill, who accepted 8-bit log for his purposes, 12-bit linear is "prodigiously wasteful". But if you don't accept the visual quality of 8-bit log, then 12-bit linear is simply inadequate.
Your original question concerning 12-bit log vs. 12-bit linear raw output is moot, since these sensors can't provide 12-bit log.

Dennis Couzin
Berlin, Germany

>I should have qualified my answer to non-VFX works. Isn't this the majority of video?

There are many cases where people shoot to RED because they are shooting for green screen or theatrical release. RED is great for VFX for HD work, because it both oversamples and does not chroma subsample. Also, RED provides tremendous amount of leverage for color manipulation, provided your cam operator shot with his eyes open. But for quick turnaround events or newscasting, where you need speed and storage efficiency over quality, there is less reason why you should shoot to RED, unless you have a RED Rocket handy.


>Your original question concerning 12-bit log vs. 12-bit linear raw output is moot, since these sensors
>can't provide 12-bit log.

Yea, that's what I quickly realized. Storing sensor data in 12 bit lin with good compression would be the best solution if the data came from a 12 bit AD converter. That or 10 bit log. Not sure how storing in log will affect matrix based color processing, as I was under the impression that storing data in log was superior to lin, given both are of the same bit depth. (surely you'll do a log2lin conversion in floating point before you apply the transformation)



www.strypesinpost.com

Also, it's worth remembering that the image processing is done in floating point and 16bit linear, and hence it's good to keep the bit precision in a rendered file higher than you'd actually need for strict 12bit linear. The difference isn't always utterly obvious, but in it is visible in some processing conditions.

Bit depth also gets confusing when you get to lossy codecs. Lossy codecs generally discard "noise" as part of their quantizing process, and it's noise (or dither) that is necessary to avoid banding above and beyond having a good bit depth.

Graeme

Good point that a lossy codec might discard dither as noise and so reduce the effective bit depth. I'm curious: where is dither typically introduced in the image chain? Do some cameras output dithered images? Is dithering just spatial or also temporal? Does system noise sometimes accomplish useful dither?

System noise is usually the dither in the system. It's somewhat computationally expensive in a camera to add dither, and probably more than worthless if the signal is directly followed by a lossy codec. For imaging, the dither is usually spatial.

Graeme