Is R3D 4:4:4 or 4:2:2?

[Adrian Jebef]

Bayer has been suggested as good for 4:2:2, but in reality it is just a different beast that does not even match 4:2:0, as not every pixel is green, but it canbe made to do a good job against 4:2:2, and even 4:4:4 (just an opinion) given suitable conditions of course. But in reality it is close to 4:4:4 (real color gaps sub pixel) when resolution is dropped in half or so from debayered etc. So, some might use this to claim that real color resolution is half etc etc, but let's not go there.

Why not go there? That's what makes REDUSER so much darn fun...

 

[Wayne Morellini]

Because it is just a pain, and with debayering, not such an issue. See, in nature color tends to change less than the brightness component. So color tends to change less pixel by pixel, and given some other cues, including overlap between the channels response curves, debayering routines can do a good job at guesstimating the color component per pixel.

I just realised something, it is possible to see purer colors. Because if overlap in the response curves for colors in our eyes, individual primaries are often picked up by multiple receptors in the eye reducing purity, it should be possible to either pick areas that have no cross response, or trick the other receptors in not seeing a color. I can't remember if this is possible, but interesting.

Like the dual color blind inspirational reforming political leader might say , "There are no white people ,.. there are no yellow, red or black people in this world,... there are only all us green people..!"

 

[Peter Dmitriyev]

So what does that mean to us layman who are only viewing the final produced image with our eyes and not intellectually as a series of numbers? Does this mean according to one of the last posts that the Sony F35 which does true 4:4:4 has basically greater and truer color fidelity than Red MX/Epic etc? Does this mean it's colors are truer, more accurate, richer, what? Does this explain Red's micro-tone separation problems on skin tones etC? Just wondering.

 

[Noah Yuan-Vogel]

I hardly think that the F35's particular color filter array makes it significantly more "true" 4:4:4 than something with a bayer pattern, it is still an array and any one photosite still only has one color filter on it. From what I understand, the C300 does a similar thing with its Bayer CFA in that it treats quad pixels as RGB pixels, R/G/B are still not co-sited. I would tend to think that these cameras that skip the demosaic/debayer step aren't giving you more color information, they just give you less luminance information since they do not attempt to reconstruct or interpolate missing RGB data from the full pixel array. So maybe the F35 is 4:4:4 in 1080p terms, but using that same terminology in 4k terms, the F35 delivers 4k 2:2:0, whereas RED delivers 4k 4:2:0 (sorry Graeme for perpetuating the misuse of 4:4:4 etc outside of its intended YUV sampling realm, but for simplicity rather than accuracy it seems to be something people have an easier time understanding for comparison). Perhaps all you'd have to do to get back F35-like 4:4:4 from RED 4kHD is do a nearest-neighbor downscale to 1080p? Or maybe RED will eventually release their 1080p RGB recording mode mentioned years back, and maybe it will just read quad pixels as RGB pixels without debayering, but why would we want that over 4k debayered and downscaled?

Maybe what we need are displays that use the same CFA as our cameras, perhaps OLPFs and aliasing and debayering would no longer have to be so problematic. Actually it just occurred to me that that is basically what the F35 does since most LCDs use a similar RGB stripe arrangement for their pixels...

 

[Graeme Nattress]

As I mentioned above, what you're talking about now is an complete abuse of terminology. 4:2:2 etc. refer to Y'CbCr data and have nothing to say on measured resolution. If you want to talk measured resolution you're going to have to be very careful on your terminology....

F35 doesn't measure full 1920 horizontally - it's soft. That's why it's not a 3k camera (or 6k as others have claimed camera). Vertically you get pure luma aliasing, horizontally it's pure chroma moire. The chroma moire is because there's 1/3 output pixel offset in the red and blue compared to green, so you get a wave of rainbow through the image of what should be black and white! The luma is due to inadequate optical filtering.

So there's your "True 4:4:4" - go on - take it, warts and all. As I said above basic sampling theory denies it to you.... And if you're tempted to go the RED route of higher resolution acquisition and "do a Canon" and deny you access to that higher resolution data, you won't be totally happy either because the "average green" demosaic is softer than a full demosaic (or downsampling demosaic) and still shows chroma moire which is not visible in the full demosaic from the same data (because I coded up a such a decode to try it out...) I'd actually say Canon's approach is better than that of Sony's on the F35 though...

Graeme

 

[Wayne Morellini]

So what does that mean to us layman who are only viewing the final produced image with our eyes and not intellectually as a series of numbers? Does this mean according to one of the last posts that the Sony F35 which does true 4:4:4 has basically greater and truer color fidelity than Red MX/Epic etc? Does this mean it's colors are truer, more accurate, richer, what? Does this explain Red's micro-tone separation problems on skin tones etC? Just wondering.

That Bayer is adequate if you treat it right, especially downsized right, but not as good as triple sampling. My pixel shift handled right might be suitable, but less accurate again, at 8k+ it maybe not much of an issue. I have no preoccupation or interest in promoting Bayer. There are so many ways to slice things up, one way often use to discredit something is failing to consider best hadeling methods. People are ussually so preoccupied with their field they don't consider everything. I think, somewhere I figured something out about the Sony diagonal pattern, that it is a sort of pixel shifting arrangement that can be processed to give the 8k, I think I might have thought about a different but related scheeme 8 years ago. A square pixel cuts out four corners of diagonal pixels, in between these four pixels you have another four pixel cirners overlapping, but if you look at it another way again, you have pixels with largerly one diagonal pixel, and little corners. A demosica's (is that the term ) delight that can yeild a calculated 8k, or less with good enough accuracy and removal of many problems with the unprocessed image. Interesting, but I would like to see the stats on it. When display pixels and recording pixeks are at the same orientation, you can't do this, no corners, 4:4:4 or whatever you may resort to resolution reducing tricks to get rid of jaggies, that this maybe trying to archive as part of the rotation. I wonder what the real reason is, could it be something ingenious as that. Maybe they did an interview with that Japanese industry site, "asci" or something.

 

[Adrian Jebef]

Graeme:

This thread was not started to discuss the use of proper terminology. The truth is that our industry's terminology is not correct or incorrect but simply in a constant state of flux. 4:4:4 was coined when we may have both been in grade school. Technically, the term implys a specific technical conversation. However, in 2012 the actual importance of full RGB color sampling (4:4:4) has changed relative to emerging thechnology like RED's R3D file format. So the question still begs to be answered: How does the machine work? Does the R3D contain full luma resolution and full RGB color resolution (fidelity)?

Wayne is on target here when he says "Bayer is adequate if you treat it right, especially downsized right, but not as good as triple sampling. My pixel shift handled right might be suitable, but less accurate again, at 8k+ it may not be much of an issue." Summed up any Bayer-type CMOS sensor can get close to proper 4:4:4 full color sampling by over sampling luma and chroma resolution with the intention of downsizing to an intended target resolution. Sample at 4K and conform to 2K/HD. The ARRI Alexa, Sony F65, and Canon C300 work this way. The Epic is designed to conform for max luminance resolution regardless of color. The debayering and software conversation process can be tweaked to make color reproduction seem more complete but it can never add discreet RGB information that was never sampled at the sensor to begin with.

 

[Graeme Nattress]

"Does the R3D contain full luma resolution and full RGB color resolution (fidelity)?" It contains the full detail in both luma and chroma that the sensor saw.

I'm sorry, but questions about Y'CbCr chroma subsampling when asked of sensors is like asking what hair colour is bald. It's the wrong question to ask - it's meaningless. The correct question to ask is, in the final RGB image that is decoded from the raw R3D files, what is the measured luma and chroma resolution. That question has been answered above.

Graeme

 

[Adrian Jebef]

G:

Yes the sensor can only record what it can see. And in the Epic's case it is only seeing half of the Blue and Red component to its intended 4K RGB image. Does the Epic's R3D file contain the full luma and chroma resolution that was sampled at the sensor? Yes. Is that resolution 100% luma, 100% Green, 100% Red, and 100% Blue per pixel at 4K? Because of the inherent design of the Epic's Bayer sensor the answer is no.

Chroma subsampling and baldness aside, let's not decide that these issues are far to "technical" to even discuss with those who actually shoot with these cameras.

 

[Graeme Nattress]

Adrian, even taking a theoretically perfect RGB sampling sensor, it's still not going to have 100% of each of R,G,B because it can't avoid sampling theory and still needs an optical filter. As I mentioned above, Sony tried to do RGB sampling with the F35 and it just lead to chroma moire and luma aliasing. Foveon has poor colorimetry, noise issues (from the layered sensor) and luma aliasing (from lack of optical filtering). Three chip cameras are much lower resolution, and don't work well with large sensors or cinema glass.

The only real solution for the problem is exactly what we're doing - high resolution sensors, optimum optical filtering.

These issues are not too technical - that's why we discuss them.

Graeme

 

[Graeme Nattress]

Adrian: "Actually the ARRI Alexa utilizes a 3K+ Bayer sensor which downscales to 2K/HD for proper 4:4:4 color resolution.

If you'd like to join the discussion check out the "Is R3D 4:4:4 or 4:2:2?" thread."

The sensor area that is read is 2880 pixels across, so that's sub 3k. In camera demosaic is quite "rough" indeed, leaving masses of chroma moire that makes it through quite clearly to the prores recording. Measured luma resolution is low, with limiting resolution around 1600 or so (out of 1920), but it's aliasy out there. MTF50 is sub 1k. Canon C300 does better with lower chroma moire, MTF50 about 1200 or so and limiting resolution is full 1920, but has corresponding aliases to go with that figure - the luma aliasing is more obvious. MTF of the F35 is similar to the Alexa - those extra pixels on the sensor didn't do it any good in terms of MTF of aliasing, and it has very different characteristics horizontally than vertically which I don't think is optimum for a digital cinema camera.

So, your statement about the Alexa is quite incorrect.

Graeme

 

[Adrian Jebef]

G:

Again, though you may be technically correct you are missing the point of this thread. Black and white resolution is a given. Color resolution is not. The Alexa may have lower MTF and Arriraw may not be 3K+ but it does sample color much more precisely than the Epic for its intended target resolution.

Sharpness or perceived resolution is an entirely different conversation. I'll leave the question of which camera is sharper up to the Cinematographers and not the Technicians. The point I'm making is that Bayer-pattern sensors provide full (or almost full) color reproduction only when they over sample (4K) and then downsize to their intended resolution (2K/HD).

 

[Graeme Nattress]

"Arriraw may not be 3K+ " - so why did you say it was?

"but it does sample color much more precisely than the Epic for its intended target resolution." but it's "intended target resolution" is less than 1/2 that of Epic's, so in absolute terms it's producing an image to fill the screen from 2880 across rather than 5120 across. You can take any camera no matter how poor and have a target resolution of an 8x8 square and get utterly superb colour resolution, but you've still only got an 8x8 square. Think in terms of what you fill the screen with. Anyway, with 1440 red / blue pixels across, it still doesn't have enough to make a fully sampled 1080p in the chroma.

" I'll leave the question of which camera is sharper up to the Cinematographers and not the Technicians." - why would you do that? MTFs are easily measured and interpreted.

I fully understand the point you're making, but it applies to all sensor types due to basic sampling theory, that optical filters don't have sharp cut-offs and that we want aliasing and moire minimized.

Graeme

 

[Adrian Jebef]

G:

Agreed. But again I like to frame these discussions in the creative realm (cinematographer) rather than the technical one (DIT). Some find the Alexa to have subpar 2K sharpness because of some of the issues you've raised. Others find those same attributes appealing because they directly influence the cameras perceptual resolution. Which some believe gives it that "smooth, organic, non-HD look" (that some DPs prefer). I personally prefer the Alexa's rendition of color resolution over the Epic's. Especially when dealing with skin tones. I understand that RED's philosophy is technical resolution first and foremost. I simply feel that color fidelity and dynamic range are more important in a 2K+ motion picture camera system than luma resolution.

So how should we wrap this up? You can derive a full-resolution RGB image from a R3D file but that image will not have fully sampled, complete Red and Blue information at 4K. ?

 

[Stuart English]

I can't answer for Graeme, but this is what I have read :


I have an idea about you / you product, and I want you to tell me why you / your product act that way / work that way.

You in turn explain to me that the idea I have is incorrect, and you give me a detailed technical explanation of why that is.

I tell you that irrespective of the accuracy or merits of your detailed technical explanation, my idea is still correct.

And then I'd like you to agree that my idea is correct and my re-expression of your technical explanation is also correct.



So no, I don't think this discussion can be wrapped up quite that way.

Unless the wrap up includes the comment that each person has their own point of view and we respectfully agree to disagree.

Which is perfectly O.K.

 

[Graeme Nattress]

Adrian, the Alexa has a built in low-contrast filter as part of the OLPF/IR filter stack. If you like that particular look a quarter soft or similar will give a close approximation for any other camera.

As for colour rendition - the above technical discussion doesn't support your conclusion as you are thinking in terms of HD resolution rather than "what you see on the screen" - the image from the sensor has to fill the screen, and the Epic puts what, 3 times more pixels on the screen.

Technical performance is vital - but it's got to be that the camera "gets out of the way" of the image. The image is the domain of the cinematographer and creative team. We don't imply a rendering intent - like the soft low con look as a "built in". We believe the camera should be neutral to the intent of the creative user and allow them to define the look.

We like resolution because it allows us to properly filter the image optically leading to a very low level of aliasing artifacts. Film doesn't have aliasing artifacts so we need to emulate that - it's the right thing to do. But if we had a low resolution sensor and properly optically filtered we'd have way too soft an image. Our MTF would be terrible leading to a miserable image. So we go high resolution, low aliasing - both traditional film qualities.

Graeme

 

[Nick Morrison]

Adrian, the Alexa has a built in low-contrast filter as part of the OLPF/IR filter stack. If you like that particular look a quarter soft or similar will give a close approximation for any other camera.

As for colour rendition - the above technical discussion doesn't support your conclusion as you are thinking in terms of HD resolution rather than "what you see on the screen" - the image from the sensor has to fill the screen, and the Epic puts what, 3 times more pixels on the screen.

Technical performance is vital - but it's got to be that the camera "gets out of the way" of the image. The image is the domain of the cinematographer and creative team. We don't imply a rendering intent - like the soft low con look as a "built in". We believe the camera should be neutral to the intent of the creative user and allow them to define the look.

We like resolution because it allows us to properly filter the image optically leading to a very low level of aliasing artifacts. Film doesn't have aliasing artifacts so we need to emulate that - it's the right thing to do. But if we had a low resolution sensor and properly optically filtered we'd have way too soft an image. Our MTF would be terrible leading to a miserable image. So we go high resolution, low aliasing - both traditional film qualities.

Graeme

One of my favorite descriptions of the red approach and methodology. Red's scientific, technical attempt to replicate the almost indiscernible, organic qualities of film. Honoring all the things we love about film, and replicating them in a digital medium that opens up worlds of additional, wonderful possibilities....

 

[Adrian Jebef]

Stuart:

"Is this to be an empathy test?"

This thread is discussing the technical and perceptional ability of Bayer-type sensors to properly sample RGB color per a specific target resolution. I'm not sure exactly what your comments refer to. But if the question you are asking is: "Can we all just get along and grab a beer together?" The answer is a most resounding "Yes"!

 

[Stuart English]

I see no discussion, that's what my comment refers to.

Where this thread goes next is up to you. If it's a statement that we shall just have to agree to disagree, we can respect that.

 

[Adrian Jebef]

As for colour rendition - the above technical discussion doesn't support your conclusion as you are thinking in terms of HD resolution rather than "what you see on the screen" - the image from the sensor has to fill the screen, and the Epic puts what, 3 times more pixels on the screen.

Technical performance is vital - but it's got to be that the camera "gets out of the way" of the image. The image is the domain of the cinematographer and creative team. We don't imply a rendering intent - like the soft low con look as a "built in". We believe the camera should be neutral to the intent of the creative user and allow them to define the look.

We like resolution because it allows us to properly filter the image optically leading to a very low level of aliasing artifacts. Film doesn't have aliasing artifacts so we need to emulate that - it's the right thing to do. But if we had a low resolution sensor and properly optically filtered we'd have way too soft an image. Our MTF would be terrible leading to a miserable image. So we go high resolution, low aliasing - both traditional film qualities.

Graeme

The vital difference is that film has 3 complete color dye layers in which to render all the colors of the rainbow. The Epic has one complete Green "layer" and two half Red and Blue "layers". And as far as putting pixels on a screen unless you have a 4K monitor or display device you are not seeing the Epic's full resolution, but rather a downscaled version of it. And yes, I am mainly thinking in terms of HD resolution but this applies to 4K as well (if and when there is a proper standard).

RED's philosophy of "getting out of the way of the image" is unique and interesting. However, the image is the thing. By not processing a proper image out of the Epic (via HD-SDI) we have no proper way to analyze or alter the image we are making from the set. I get that you like resolution. But why choose between high resolution with limited color reproduction vs. low resolution with complete color reproduction? We want is all: Super high resolution with complete color reproduction.