Two questions about downscaling 4K

[Stokestack]

I keep asking these in relevant threads, but so far no answers. I'd like to have one set of lenses and have the 35mm DOF regardless of what my stored image size is. So...

1. Is it technically possible to scale raw images (hopefully in camera) and keep them "raw" (in other words, keep their raw latitude)? If there are R, G, and B photosites on the sensor, isn't it possible to resample and merge their values just as you would with pixels in a non-raw image? Then store them in a file that looks (in terms of data layout) like a 2K raw file?

2. Why is the RGB storage only 4:2:2? This becomes a bigger deal if the answer to #1 is no, leaving 1080P RGB our only 4K-to-smaller scaling option.

Thanks!

[Rob Lohman]

1. this is not a feature for now. Who knows what will happen in the future with some Graeme magic, but not right now. Shoot 4K and downsample that or shoot 1920 x 1080 RGB which is fairly close to 2K.

2. where did you get this? RGB is "4:4:4", there is no real "4:2:2" format for RGB.

So when you shoot 1920 x 1080 or 1280 x 720 in RGB you will get the full resolution and full sampling!

[Jim Arthurs]

So when you shoot 1920 x 1080 or 1280 x 720 in RGB you will get the full resolution and full sampling!

One important thing to remember for those who are only dreaming of mega-rez 4K... the 1280 by 720 (the other HD) will be SPECTACULAR. Without question, the best in class. And plenty clients want it.

Remember, this is not DVCPRO 100 720 which is an anemic 960 by 720, or exactly 2X SD resolution.

So, since 720p is a viable format, RED will give the worlds best 720p... full raster, and massively downsampled in a bullet-proof 4:4:4. There's plenty of life and demand for 720p, and just think of the long record times on one RED drive...

[Chris Kenny]

Yes, there's going to be a big difference between a lot of the 720p out there now (4:2:2, DCT compressed, 8-bit, 960x720, often created using barely-HD sensors, etc.) and what RED will produce (4:4:4, wavelet compressed, 12-bit, 1280x720, and massively oversampled).

[GlennChan]

Other cameras actually have higher bit-depth A-->Ds behind them than Red (i.e. 14-bit). Compare that to Red's 12-bit A-->D. So in particular numbers, other cameras might have an edge over Red.

2- Alan Roberts has an intriguing post over at dvdoctor.net, claiming the HDX900 has 12 *measured* stops of exposure range. And he's credible, doesn't work for Panny, and definitely knows his stuff. 12 stops is slightly greater than Red (although Red's exposure range might go up once you downsample??????).

http://forums.dvdoctor.net/showthread.php?p=279767&highlight=stops#post279767

Not saying that it's necessarily better, but definitely interesting.

[Laco Zamba]

Compare that to Red's 12-bit A-->D

Is it true?
I don't think so.

[Jeff Kilgroe]

And since we're on the subject of best in class 720p...

Can someone clarify for me, is 720p downsampled from the full 4K sensor size like 1080p is? I would assume it is, so we get the benefits of the full sensor size and DOF. Also is it possible to shoot 720p using a windowed sensor area so we can have frame rates in excess of 60fps? In other words, could the 2K windowed sensor area be used to record 720p @ 120fps on camera?

I've seen such things mentioned in various threads here, but have not seen actual clarification of this. 1080p downsampled from the full 4K sensor size and recorded as RGB on camera is probably what I will use for most of my real-world work for clients. It just fits like a glove and a lot of them are starting to ask about BluRay/HD-DVD. But if I could also shoot a windowed 2K area at up to 120fps and record it downsampled to 720p RGB, that would be great.

[Jeff Kilgroe]

OK, I'm assuming what I asked about can all be done based on the format options chart at RED.COM. So correct me if I'm wrong.

[Rob Lohman]

Yes, that should be how it works Jeff.

[Chris Gearhart]

Other cameras actually have higher bit-depth A-->Ds behind them than Red (i.e. 14-bit).

As for 14 bit color depth of other cameras--and I am not up on what cameras are 14, but just taking your statement--I wonder at what cost, SNR-wise? Is their sensor really capable of pushing that much latitude? And if they are not getting 13 or 14 stops, what is the benefit to storing/transmitting that much data? I wonder if that isn't a spec in search of a benefit. I am speaking from complete ignorance, however. :weight_lift:


12 stops is slightly greater than Red

I wouldn't count on that. Graeme has been getting wacky-compulsive in getting the full 12 stops out of the 12 bit space. His posts are very . . . suggestive. As always, everything is subject to change, count on it (including this sentence).

[Jim Arthurs]

Other cameras actually have higher bit-depth A-->Ds behind them than Red (i.e. 14-bit). Compare that to Red's 12-bit A-->D. So in particular numbers, other cameras might have an edge over Red.

I think the RED's massive over-sampling and true 4:4:4 will "win" over any other camera and specs, dubious dynamic range claims aside. Easy to test this theory; take a digital still camera, shoot the same subject as your HVX/Varicam, downsample the still image to equal size and compare.

Remember, this sharpness will be real and created from the downsizing of a non-sharpened source many many times larger than the final target, and not provided by edge enhancement to a native 720p image.

While most people are ga-ga (and rightly so) over the potential of working with 4K, I am looking at what my clients want/need today and the reality is 720p or 1080 will suit them just fine. And, due to Varicam and HVX... they've never even SEEN real 720p played back at full raster... so the RED is an upgrade without even changing formats.

Oh, and 720p at 120 fps might have some attraction as well :)

[Michael Schrengohst]

I think the RED's massive over-sampling and true 4:4:4 will "win" over any other camera and specs, dubious dynamic range claims aside. Easy to test this theory; take a digital still camera, shoot the same subject as your HVX/Varicam, downsample the still image to equal size and compare.

Remember, this sharpness will be real and created from the downsizing of a non-sharpened source many many times larger than the final target, and not provided by edge enhancement to a native 720p image.

While most people are ga-ga (and rightly so) over the potential of working with 4K, I am looking at what my clients want/need today and the reality is 720p or 1080 will suit them just fine. And, due to Varicam and HVX... they've never even SEEN real 720p played back at full raster... so the RED is an upgrade without even changing formats.

Oh, and 720p at 120 fps might have some attraction as well :)

I agree 110%....I only shoot timelapses with my Nikon D200....
I had a client with an HVX200 and we shot the same timelapse
of a property...going day to night.....
The Nikon image blows the HVX out of the water....I am hoping
that the RED can even improve on the Nikon D200 image quality.

[Stokestack]

2. where did you get this? RGB is "4:4:4", there is no real "4:2:2" format for RGB.

This has come up several times in the forums. Even today, the Red site says that the only onboard downscaled-from-35mm format is Redcode 1080P or 720P (http://red.com/formatoptions.htm). And on the Technical page (http://red.com/techspecs.htm) under Redcode, all you see for 1080P or 720P is this:

10 bit 4:2:2 1080p / 1080i / 720p

So if that's not true, it's good news. But maybe this info on the Web site should be revised.

[Nick Shaw]

Yes those sections of the RED site out of date. A lot of it has not been updated for a while. Graeme has said everything is RGB 4:4:4 or RAW. 4:2:2 will not be a recording option, only available on the HD-SDI outputs. (All specs subject to change etc.)

[GlennChan]

As for 14 bit color depth of other cameras--and I am not up on what cameras are 14, but just taking your statement--I wonder at what cost, SNR-wise?

Hmm it might just be the latest Panasonic cameras that have 14-bit A/D. From what I know, higher bit depth analog-digital convertors bottleneck the maximum dynamic range possible (i.e. Red will never have more than 12 stops). By having better A/D, you raise the limit as to how much dynamic range is possible.


Is their sensor really capable of pushing that much latitude?
According to Alan Roberts, yes.


And if they are not getting 13 or 14 stops, what is the benefit to storing/transmitting that much data?
The camera applies gamma correction/compression (like every other normal video camera). You basically squash all the data to fit into a smaller bit depth.

Also, our visual response roughly follows a power function (roughly 3.0 if you agree with CIE L*; Graeme quotes a different figure). Or stated differently, we are more sensitive to shadow detail than to highlight detail (if you look at the intensities in their linear form). So you can squash the highlights a lot (and not so much the shadows), and the signal will fit into 8 bits (or 10 bits).

This is what video cameras have been doing for a long time.

2- Most cameras don't apply straight gamma compression/correction and they don't follow Rec. 601 or Rec. 709 to the letter. They almost always have knee settings, and they tend to have different gamma curves or custom transfer functions. They give you different options in how you want to map the sensor output (i.e. however many stops of dynamic range the sensor delivers) into what you see.

3- A straight "stops of dynamic range" figure may not be entirely indicative of actual stops of dynamic range.

In most cameras, the highlights tend to shift colors. Wacky-looking highlights just aren't that useful. You might want cameras that corrects this for you (it looks like Red is doing this?).

[Rob Lohman]

Stokestack: thanks for the link, that's indeed incorrect. I'll see if I can get that changed.

[Stokestack]

No problem. Glad to hear it's incorrect. The scaling option now seems much more appealing!

[Andrew M.]

[B]2- Most cameras don't apply straight gamma compression/correction and they don't follow Rec. 601 or Rec. 709 to the letter. They almost always have knee settings, and they tend to have different gamma curves or custom transfer functions. They give you different options in how you want to map the sensor output (i.e. however many stops of dynamic range the sensor delivers) into what you see.


But if cameras do apply gamma correction, is this done right in the hardware on the CMOS chip before it is converted from analog to digital or only after the A/D conversion is done on linear signal?

Andrew

[Graeme Nattress]

Well, the specs of the HDX900 are 54 dB (typical), which to me is no more than 9 stops. 12 stops does imply at least a 72db SNR.

Graeme

[Graeme Nattress]

And yes, proper downsampling to HD from 4k should drop the noise floor and give you more dynamic range.

Graeme

[tj williams]

14 bit color depth is offered on the new Panasonic HDX900 we shot one of these off against an F900 (which is I believe 10bit) a while back and the Panny pasted the Sony! which featured much higher cost. How much of an issue is color depth differences between 10bit 12bit 14bit is there any real world description of possible differences?

[Graeme Nattress]

Well, A-to-D bit depth governs how many code levels are used to describe brightness. This relates to dynamic range and to noise levels. You don't want to be in the situation where your a-to-d quantisation noise is greater than your sensor noise, as then you'll loose information and dynamic range. But similarly, there's not much point having many more bits (ie smaller quantisation noise) than the noise floor of the sensor. It's best to record 1 bit of noise, but I don't see much advantage in recording more than that.

So, what would you see - would mostly be in noise levels and dynamic range.