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

[Adam Beck]

Television programs generally don't go through "rigorous" post production, and now that Next Element is no more, I don't really know of any Red shows being finished directly from the R3D files. If I'm wrong on that, please feel free to correct me.

Television and features are very different beasts, with very different requirements, conditions, post schedules, and post production approaches. The advantages of Red are very clear in features, as they largely relate to resolution and its ability to hold up under very large screen viewing conditions, as well as flexibility when the original files are made available, which is more common on features and more necessary due to the use of "look creation" techniques. They're not so clear in television, where speed and simplicity are the orders of the day, and where the loop is not nearly as closed in terms of people doing what they're supposed to do all along the post path.

Mike,

Would you say that Redgamma 3 could potential increase the amount of Red's on television productions?

 

[M Most]

Mike,

Would you say that Redgamma 3 could potential increase the amount of Red's on television productions?

Honestly, no. Among other things, I've yet to see an image processed through Redcolor3 and Redgamma 3 that includes a human face in closeup. But more to the point, I don't think color improvements are the overriding factor here. Alexa images are generally delivered as LogC because that allows for a lot of detail retention and a lot of flexibility in grading, but with a reasonably good starting point provided by the standard Arri LUT with a color matrix. Colorists in general - and many cameramen, too - want images they can control, but that get to a good starting point quickly. That's what RedlogFilm and Redcolor2 already provide. The only real advantage of using something like what we've been shown in test images is if one wanted to forego color grading altogether and accept the baked in color as the final product. That is something that largely went away with cameras like the F900 and other earlier generation cameras based on video camera designs. In the network level television world, I don't see that as either desirable or likely. So while Redcolor3 will likely present improvements over Redcolor 2 (which was considered a bit too low on saturation by many colorists, although a bit more pleasing than Redcolor), I don't see wider acceptance based on the new color science alone.

Now, if an HD image could be recorded on camera using Redcolor 3 and RedlogFilm, well........

 

[Evin Grant]

Now, if an HD image could be recorded on camera using Redcolor 3 and RedlogFilm, well........

This I certainly agree with. Also as a personal preference I tend to like less sturation. The default 709 space on the Alexa looks much too pumped and sort of circusy to my eyes. Of course this is easily corrected for with the Log-C but that may account for my preference of Epic skin tones.

 

[Adrian Jebef]

I think the simplest definition someone wants is "1920x1080 resolution in all 3 colors." RED has enough sensor sites to satisfy that requirement for 4:4:4.

If you view 4:4:4 as a fancy way of saying 1:1:1 (e.g. as a ratio) then no the ratio is not 1:1 for all channels of resolution @ 4k. Once you supersample though down to HD (and compare to Alexa) it's *at least* 1:1.

And none of that has anything to do with skintones. That's a colorimetry question not a resolution question. If it were a resolution question then SD skintones from RED would be better than the same image in HD--which isn't the case. In fact I would say the opposite is true. At HD you can dither the color and get perceptual color accuracy beyond that of the same bit-depth in standard definition. I still see agree that skintones out of RED are more frustrating than out of some other cameras and film in particular--hopefully something that Redcolor 3 will resolve.

Gavin:

I am not looking for anyone here to "give me the answer I want". I am simply interested in provoking a good disscussion about things everyone seems to understand completely yet constantly offers up different definitions by way of zero explanation.

4:4:4 is a ratio and fundemantly refers to 1:1:1 or equal RGB sampling. And as you say the Epic nor its R3D file contains 4:4:4 color information at 4K. Unfortunately, the RGB image from the Epic is only decoded at its native resolution (4K). From there it can be resized to 2K/HD but it is not re-debayered for a proper HD 4:4:4 output by using adjacent RGBG pixels to calculate finer color rendition.

It would be great to have this option though...

 

[Graeme Nattress]

"not re-debayered for a proper HD 4:4:4 output by using adjacent RGBG pixels to calculate finer color rendition" - I'm sorry, but with statements like that you really don't know what you're talking about.

Graeme

 

[Adrian Jebef]

"not re-debayered for a proper HD 4:4:4 output by using adjacent RGBG pixels to calculate finer color rendition" - I'm sorry, but with statements like that you really don't know what you're talking about.

Graeme

G:

So please tell us. I do not work for RED. How is downscaling handled with an R3D file targeted to 2K/HD?

 

[Graeme Nattress]

Well, you've told us "but it is not", so to say such a thing you must know "what it is"? And what is "re-debayered "?

To answer your question, typically a full-demosaic is performed followed by a downsampling process. We provide for a number of selectable downsampling filters, all of which use adjacent pixels.

Graeme

 

[Les Dittert]

Debayer native sensor data.
Resize image to 2K or whatever using selection of methods.
The smaller you make the output image the closer you get to what is being referred to as 444.

G:

So please tell us. I do not work for RED. How is downscaling handled with an R3D file targeted to 2K/HD?

 

[Adrian Jebef]

Debayer native sensor data.
Resize image to 2K or whatever using selection of methods.
The smaller you make the output image the closer you get to what is being referred to as 444.

But color accuracy depends entirely on the RGB photosensors on the camera and the chosen debayering method. A debayering to 4K would yield different color results than a debayering to 2K/HD. A debayering to 4K that is then resized to 2K/HD will deliver the same color results.

 

[Adrian Jebef]

Well, you've told us "but it is not", so to say such a thing you must know "what it is"? And what is "re-debayered "?

To answer your question, typically a full-demosaic is performed followed by a downsampling process. We provide for a number of selectable downsampling filters, all of which use adjacent pixels.

Graeme

Hmmm... I was hoping for a more in depth explanation.

So can you tell us where the selectable downsampling filters are located in RedCine X?

 

[Graeme Nattress]

Adrian, again you're not making much sense. The colorimetry of the camera is independent of demoniac and downsampling methods applied.

Graeme

 

[Graeme Nattress]

Hmmm... I was hoping for a more in depth explanation.

So can you tell us where the selectable downsampling filters are located in RedCine X?

When you set up an export preset, in the "options" tab you'll find your choice of downsampling filter, ranging from sharpest through to softest from sinc to wide gaussian.

Graeme

 

[Adrian Jebef]

Adrian, again you're not making much sense. The colorimetry of the camera is independent of demoniac and downsampling methods applied.

Graeme

OK. So when I find less than appealing skin tones from the Epic is that the result of the camera's colorimetry and not the RGB color fidelity issues inherent in a Bayer sensor?

Are you suggesting that over sampling chroma for a given target resolution (8K sample/4K target) creates has no discernible difference in a camera's color fidelity?

 

[Evin Grant]

Debayer native sensor data.
Resize image to 2K or whatever using selection of methods.
The smaller you make the output image the closer you get to what is being referred to as 444.

What does not make sense here is that you seem to be implying than a downsampled image will somehow have more color fidelity than the original, but if you look at a Red 4K image at full resolution and one downscaled to 2K there is no preceptible difference in color.

OK. So when I find less than appealing skin tones from the Epic is that the result of the camera's colorimetry and not the RGB color fidelity issues inherent in a Bayer sensor?

Sort of, it's your opinion of the color you're seeing out of your particular viewing system or monitor on set.
I have the same problem with the Alexa 709 space, it looks too saturated and funky, but the s-log has the info there. All you need to get the skin tone you like out of an Epic is spend a few seconds in RCX.

BTW the Alexa uses exactly the same bayer pattern filter design, so if you like Alexa colors then there is no argument that the filter pattern is to blame.

 

[Adrian Jebef]

What does not make sense here is that you seem to be implying than a downsampled image will somehow have more color fidelity than the original, but if you look at a Red 4K image at full resolution and one downscaled to 2K there is no preceptible difference in color.

Which is exactly my point. The Epic does not see color like I see color by eye. The Alexa (for example) seems to mostly come dangerously close. We know the Epic has WAY more resolution in measurable luma and chroma. So what gives? If color stays the same at 4K and 2K/HD why build a 3K sensor (Alexa) to deliver HD? Or why build a 8K sensor (F65) to deliver 4K?

 

[Evin Grant]

No camera can see color the way you see color, and if you happen to see color the way an Alexa sees color I suggest you visit an optomitrist. Seriously, you obviously do not like the way the Red processes color, great, fine, don't shoot with one. That does not mean that it's incapable of recording pleasing colors, flesh tones or your subjective definition full chroma resolution.

 

[Uli Plank]

If color stays the same at 4K and 2K/HD why build a 3K sensor (Alexa) to deliver HD? Or why build a 8K sensor (F65) to deliver 4K?

Count the pixels and don't read marketing papers only. The F65 has 20 mpx vs. 14 for Mysterium X. That's less than 1/3 more. It doesn't help resolution if I tilt my camera 45 degrees, whatever Sony is saying.

 

[Noah Yuan-Vogel]

Noah:

As far as I am aware you cannot skip the demosiac / debayering process. Otherwise you would not have an RGB image. There definitely are different ways in which to debayer the same RAW sensor data to varying degrees of success. The early days of the RED One and changes in RED's R3D processing are notably the best example.

To be technically correct for our Workflow Wizards and Problem Solvers, a Bayer color filter array (CFA) is made up of 50% Green photosensors, 25% Red photosensors, and 25% Blue photosensors. This is known as RGBG and can perhaps be expressed as 4:2:2:4 So should we agree that the Epic's R3D file is 4:2:2:4? That would be cool. I wanna get that printed on a t-shirt right now!

What I meant by skipping the demosaic process was process not attempting any interpolation as is generally done during demosaicing any bayer or other cfa into RGB where photosites=RGB pixels and for each pixel two colors must be interpolated.

As for 4:2:2:4, I can't agree, and why does it have 4 alpha channel samples?

I'm sorry I ever suggested that it was ok to continue with talking about the samples in color filtered digital imaging systems using the YUV sampling terminology that was intended only for that use. Clearly it Graeme and David are right and any talk of 4:2:2 etc when referring to bayer or other color filter patterns will only serve to further confuse concepts that are already too often misunderstood.

Does anyone else have any thoughts about the merits of the RGB stripe CFA in that the R G and B filtered photosites would correspond with the RGB stripe pattern of most modern display devices? I assume that is the purpose for that design, but any idea if that does mitigate any of the color fringing or aliasing concerns? To me it is an interesting question as to whether 4:4:4 implies that the samples are supposed to appear co-sited when in fact the signal is being sent to output devices where RGB pixels are not co-sited themyselves. This is the same reason most computer systems apply filtering to fonts and other graphical objects to compensate for the fringing that would occur from computer graphics being designed with co-sited colors but color displays mostly using RGB stripe pixels.

 

[Les Dittert]

Some of you guys are mixing up color spectral response as defined by the sensor RGBG filter vs the color spacial resolution ( ie how detailed the color information is , how 'fuzzy' the chroma is ) based on debayer method.
Skin tone issues would be blamed on the former.

 

[Wayne Morellini]

Crash, re-write. I think that this discussion has being disingenuous from many sides. If a person is relying on the wrong term, knowledge etc. Let them know what they really mean and answer based on what they mean, rather than playing around, it is extremely disrespectful, unwise, rude and stupid. Etc.

This talk about imperfect sensor recordings of reality like they are authentic reproductions in wrong. Acting like 100% fill is non authentic, 4:4:4 range spread at 100% fill done propperly, is the most accurate authentic sampling, and the only authentic sampling short of full spectrum signal recording. It is simply because it maximises the range of authentic values. Making things more like film than reality or recording reality, is not authentic. Even our eyes are not authentic at recording, but a good starting point for us.

Wayne, the Foveon white papers say that an OLPF is still necessary to reduce aliasing on their sensor. "This filter is largely unnecessary with the Foveon X3 sensor and is not used. " is opinion and not fact, and not backed up by the Foveon white papers themselves. Sure, you don't get chroma moire, but that's the easy bit to remove on a Bayer CFA. Luma moire is practically impossible to remove on any sensor type, especially on moving images.

"the fill factor alone, if near 100% effective with microlenses, can provide a significant anti-aliasing effect," - again, incorrect. A high fill factor will give the sensor a lower inherent MTF, but that's not the same as anti-aliasing.

Graeme

You miss the piont of my post. Really, has the Epic sensor pads got 100% fill not to miss ths unrecoverable detail? You know, I can find diagrams of Sony and Foveon sensor patterns, and even of the foveon sensor pad, and where the technology came from, but I haven't seen a single thing of the Red sensor and where their technoligy came from, even though Jim promised it many years back, and original container files of example frame of the technology, not a Red camera, was thought to be Canon or something like that (long time ago). At least they put out.

So, bayer sensors are likely to miss more detail then the latest foveon. As well, just because there is a primary pad at each location, and you can use debayering to fill in missing primaries, likely with some accuracy, does not mean you will not miss detail accuracy, or even whole details. A short change in other primary channels smaller than the pad often maybe missed without interframe debayering (just invented it) to pick it up as it mives relative to the pads. As well, total detail maybe missed if it is entirely in another primary channel. Put on top of this all the missed space inbetween Bayer pads. So luminance problems are still there. The way they get around it is to compromise the image by bluring it, which in practice only dies little to reproduce such detail.

As far as anti-aliasing in particular, the closer to 100% the more you pick up the luminance and color mixing of edges with the surrounding envrinment, as they pass further and further into pads. With less fill factor you piont at which you see your neughbour might be seperated missing this edge progressuin through the pads, resulting in unatural jumps from over edges frim one color/intensity to another, or what is called alaising. So, fill factor has anti-alaising benefits, more authentic then blurring.

When we turn our attention (sorry am not well very difficult to pull my head around everything I wanted to say) to foveon. 1. I doubt they have 100 fill factor through their microlens in the first place, and the extreme edges of the lens may present issues in terms of quality. 2. They could use some blurring to increase anti-alaising, but nt as much as Bayer normally. You can apply some artifical mixing from pixel to pixel through software, or use higher Res sample pixels so they overlap screen pixels. However pixel sized or bigger detail is fairly contiguous, so software can be used to calculate continuation from pixel to pixel and correct somewhat (as well as interframe debayer mentioned before). So, in this narrow respect they may have some advantage before adding filters and micro lens arrays from my understanding, for a better overall result, but still failed in other color channel issues. However, we do not need full color fidelity to make a premium quality film distribution (you know the type of grading I'm referring too) but noise is a bigger issue (which they used to use lighting for).

About the foveon white papers, when were they written and do they cover the latest architectures fillfactor?

I am sorry I cannot answer you with the accuracy and depth I would like Graeme, I am still sick, and my eyes are still watering and stinging, but things are reasonable.

The higher spatial res sensors Dragon+ will I think will embody a very key tipping point where such discussions in the future become moot and the conceptual notion of 4:4:4 becomes almost a “Dinosaur” or at least an intermediate “working theory” much like Phlogiston in the late 1600s. Six years from now no one is going to be having trouble with this

That was my point, that even with my 1/3rd pixel shift the differences become less noticable at higher resolution. So much so, that I contemplated how Bayer raw would look without debayering, Graeme?

I think 4:4:4 would still be somewhat important to special effects people Eric

Absolutely correct David, and well put.

"even if color information is captured by various layers, that doesn't mean each layer has equal resolution" - I remember the first time I looked at a film scan DPX and examined the individual channels and saw how different they can be. But it doesn't mean that film can't make a great picture, because it does! Same with RED Cameras. Although they use the often mocked (and mis-understood) Bayer Pattern CFA, they produce superb images in the right hands, and even I can make some pretty images with them too!

Graeme

Not fair, you are a Bayer genius. I should have noted your software in improving resolution etc in the last example.

However, a past problem in lens systems was that each primary range would focus at a different depth. So, if you focused on green portion red and blue could be out of focus somewhat. Also the color resolution problem on Bayer is way worse. So calculations on a well done 100% fill 4:4:4 file for color handling are either going to be simpler and or more accurate than Bayer.

F65 ?

That's the sensor I was thinking of, maybe Graeme didn't know about it to point me in the right direction than ignoring it.