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

[Evin Grant]

Debayering has also become much more sophisticated in the recent past so just because a sensor is mathematically 50% green, 25% Blue and 25% Red does not mean that this is all the usable resolution it has for each channel. We know that with a 5K Epic file we can extract 4K+ worth of measurable luma resolution so even if we use the old 4K luma numbers for the chroma we'd still have 3.6-3.8K worth of chroma resolution. Despite it's irrelevance my guess is that would equate to about 4:3.5:3.5, much closer to 4:4:4 than 4:2:2.

 

[Martin Stevens]

Debayering has also become much more sophisticated in the recent past so just because a sensor is mathematically 50% green, 25% Blue and 25% Red does not mean that this is all the usable resolution it has for each channel. We know that with a 5K Epic file we can extract 4K+ worth of measurable luma resolution so even if we use the old 4K luma numbers for the chroma we'd still have 3.6-3.8K worth of chroma resolution. Despite it's irrelevance my guess is that would equate to about 4:3.5:3.5, much closer to 4:4:4 than 4:2:2.

This math makes the most "real" sense to me.

Thanks Evin.

 

[Graeme Nattress]

I think you didn't pay attention to what Graeme said. Your statement doesn't compute.

Correct. That is an over-simplistic answer to a fundamentally complex (and more interesting) question. I prefer the more reasonable response I gave above.

Graeme

 

[Tom.Wong]

the questions shouldn't be whether it's 444 or 422. these numbers don't apply to rgb raw. the real question is, what is the full color gamut that comes off the Epic sensor. RED never released any charts on that on how wide the range of color Epic produces. RAW is RAW, but depending on the technology that goes behind it, every camera produces a different range in color gamut.

 

[Graeme Nattress]

Sensor colour gamuts are not precisely defined in the way that monitors are with their distinct primaries. Our cameras all have similarly wide gamuts, and all colour filter array cameras pretty much use the same dyes on the sensor anyway.

Where we continually try to work, improve and learn is raw development, with a strong sense of digital cinema aesthetic rather than a video aesthetic in mind. In the process of working on REDColor3 I've learned a significant number of new nuances which will influence future development.

Graeme

 

[Jonas Rejman]

Thanks for the lectures, Graeme. Always an informative read.


p.s.

And why does Graeme has time to post here, instead of finishing REDColor3?
Is it done yet? (nag nag nag nag nag)

 

[Graeme Nattress]

Gosh I hope I wasn't lecturing

Graeme

 

[Martin Stevens]

How about.....Graeme is a Genius!

I can hardly wait to experience the cinema oriented RedGamma3 etc.?

 

[Christian Jadot]

I love these threads! The one I have to read a couple times to get it to sink in.

Thanks for the detailed explinations again Graeme.

 

[Uli Plank]

More food for thought, just to show how complex things are: RED is filtering luminance (with the OLPF) quite conservatively to avoid any aliasing, but decided not to filter out all chroma aliasing, since that would soften the image too much. That's why there has been an option for chroma filtering in all the post software RED gave us. So, what about the rare cases of chroma moiré, i.e. false detail? Would you be better off having 422 or 444? I believe that depends so much on the way downsampling is done, that it's quite difficult to give a general answer.

BTW, the RED cameras have far less chroma aliasing than a Sony F3…

 

[Graeme Nattress]

The OLPF filters all in-coming light, not just luma. Although the level of filtering applied is to balance sharpness and propensity for aliasing on the luma, it also filters the chroma, albeit not to the same extent as the luma. There is the possibility of chroma aliasing, albeit at a lower level than most cameras would allow for aliasing in general, and also, remembering as you point out that chroma moire is reasonably easily filtered if it ever appears.

Graeme

 

[Jack Shanahan]

Sensor colour gamuts are not precisely defined in the way that monitors are with their distinct primaries. Our cameras all have similarly wide gamuts, and all colour filter array cameras pretty much use the same dyes on the sensor anyway.

Where we continually try to work, improve and learn is raw development, with a strong sense of digital cinema aesthetic rather than a video aesthetic in mind. In the process of working on REDColor3 I've learned a significant number of new nuances which will influence future development.

Graeme

and this is why i bought RED.

 

[Stivan Widick]

I'd pay to go to Graeme lectures. Good stuff.

 

[Les Dittert]

The OLPF filters all in-coming light, not just luma. Although the level of filtering applied is to balance sharpness and propensity for aliasing on the luma, it also filters the chroma, albeit not to the same extent as the luma. There is the possibility of chroma aliasing, albeit at a lower level than most cameras would allow for aliasing in general, and also, remembering as you point out that chroma moire is reasonably easily filtered if it ever appears.

Graeme

Hi Graeme,
Sounds like there might be a chance of some RCX debayer improvements in the near future, then ? Great !
I will soon ( possibly today ) be posting a RAW Cineform mov that will demonstrate how the Red output looks when run through that workflow. They have several methods available for debayer, switchable on the fly in firstlight.
Just an extra data point for people to see and compare!

 

[conrad gaunt]

I think you didn't pay attention to what Graeme said. Your statement doesn't compute.

I did pay attention, and it computes perfectly. I also design debayering algorithms in my spare time.. its a nonsense question, but is bayer sub-sampled colour = yes.

4:4:4 implies no colour subsampling

Yes, debayering algorithms have improved, they do a good job of guessing at missing data.. but tell someone deciphering/analysing the dead sea scrolls pixel by pixel that working from a debayered "guesstimate", compared to a 4:4:4 aligned scan, is a good thing .. but i say this tongue in cheek to make a point

its was a fun/geeky question, with no correct answer .. but ill still take mine in this particular case

 

[Graeme Nattress]

"4:4:4 implies no colour subsampling " indeed it does, but it implies that no chroma sub-sampling occurs in a Y'CbCr space as a means of compression.

"bayer sub-sampled colour = yes" indeed, but at a completely different part of the chain in a different colour space and different linearity by a different method.

Graeme

 

[conrad gaunt]

"4:4:4 implies no colour subsampling " indeed it does, but it implies that no chroma sub-sampling occurs in a Y'CbCr space as a means of compression.

"bayer sub-sampled colour = yes" indeed, but at a completely different part of the chain in a different colour space and different linearity by a different method.

Graeme

it doesn't really matter if the sub-sampling happens after a lossy conversion to yuv, for compression reasons, or whether it happens because it a bayer pattern made up of impure colour samples, since the chroma in yuv consists of R/B differences relative to green.. its similar to how many debayering algorithm work.. such as smooth hue interpolation, or any debayer that uses the green channel to guide reconstruction of the missing red and blue samples..

i agree about the colour space stuff being different etc, and that most yuv conversions are very lossy if reversed back to rgb.. and redcode doesn't do that..

So, ill try a compromise..

4:4:4 AND 4:2:2 both = reduced resolution colour (relative to luma), made worse by lossy colour space transformations which reduce R,G,B information uniformly, but leaves data that is easier to compress and transmit

RAW = similar reduced colour resolution (ratio), but not subjected to low quality realtime in-camera yuv colour space transformations= less wasteful = nothing baked in = 1/3 datarate = future proofed data (should you or someone else improve algorithms) ..

its still colour subsampled, I believe that is mentioned in the original Kodak paperwork by mr Bayer

Debayering is only half redcode success, the wavelet transformation of course, and the way it distributes quantization across the frequency range and then weaves it back together is the other big part of the equation, and the reason you have decent raw image to debayer at all,,

.. I'm new to wavelets, although I'd been using Haar functions in audio processing without realizing what I was doing.. I'm currently working on an image frequency "equalizer" ..6 band..getting there..you should have a play with wavelets if you haven't already.. you seem to have nailed debayering

 

[Graeme Nattress]

"it doesn't really matter if the sub-sampling happens..." well, yes it does because the visual result is rather different. With chroma decimation after a YCbCr transform, chroma resolution is some fixed ratio of whatever it was before it was sub-sampled. In the Bayer pattern sensor system, measured chroma resolution will be different for each colour, depending upon many factors.

Graeme

 

[conrad gaunt]

"it doesn't really matter if the sub-sampling happens..." well, yes it does because the visual result is rather different. With chroma decimation after a YCbCr transform, chroma resolution is some fixed ratio of whatever it was before it was sub-sampled. In the Bayer pattern sensor system, measured chroma resolution will be different for each colour, depending upon many factors.

Graeme

I already described some of the differences above, and why the results can be different, and in greater detail than your above post. If you throw colour information away, its subsampled. If the colour samples are reduced in resolution to start with, they're also sub-sampled

When some one asks me if is something 4:4:4 or 4:2:2, I don't assume they're comparing video standards with D.I, and the original post wasn't IMO, what they're really asking is "is the colour/chroma/whatever" subsampled. Answer, yes .

Debayering is intelligent/subjective interpolation, and bayer patterns are a form of colour sub-sampling, as are the 4:4:4 and 4:2:2 video formats

I don't think anybody will read my answers and say "so redcode is 4:2:2, or 4:4:4", because it is neither or either, depending on post decisions, formats , resolutions.. but the sensor is biased towards luma, not colour.. that is a common design. I'm not afraid to discuss differences and similarities between formats

most compression systems use prediction to encode differences, with redcode its wavelet coefficients spread across frequency/resolution bands, with traditional macro block based DCT encoding its largely the difference between G and R/B (through a conversion to YUV) and heavy blocky quantization, or temporal compression with mpeg2.. so yes, there are differences, as I already said, the biggest being the wavelet transforms instead of DCT etc IMO, ..or the fact that raw files aren't limited to in-camera circuitry with cost/ speed restrictions

..but since 4:4:4 and 4:2:2 video tends to end at 1080p 10bit, and redcode can record 5k at 12bit.. i think you should be less touchy and not restrict your answers to redcentric ones.. most people who haven't compared camera formats won't be buying Scarlets , Epics or R1s.. those that have can probably handle understanding the differences, and similarities, between signal encoding systems

A balanced discussion requires both sides of an argument, and while I disagree with little that you say, sometimes when you respond to posts you seem to take a line that is b&w and designed not to confuse people (well your busy i guess..), which is understandable, but at the expense of some accuracy.. which results in a shorter discussion and then the "magic pixies" appear

with a few of my past posts, and this one, i feel you're trying to counter me with my own arguments repackaged.. pushing against an open door

"measured chroma resolution will be different for each colour, depending upon many factors." .. well yeah, but is this a master class in stating the obvious? ..if I interpolate adjacent pixels from a 4:4:4 video source and blend that with the original colour will I not get a different colour, for each pixel?.. probably! Will it depend on many factors? .. if I make it adaptive, sure.. but explains nothing

Bayer pattern is a bandwidth reducing / colour sub-sampling design, I acknowledge many differences, but that is something it has in common with other colour video acquisition systems, including the 4:4:4 and 4:2:2 video standards. Why not start with that, then EXPLAIN the differences between that and redcode (without giving away what type of elves you use)

Sensor readout (RGB bayer pattern) -> camera colour sytem/realtime-debayer -> monitor
|
\/
Wavelet Transformation
|
\/
Compression
|
\/
Write to Media
|
\/
backup/verify ->post: inverse wavelet transformation/DSP/Colour conversions/HQ debayering etc -> 4:4:4 or 4:2:2 or DPX etc

why not explain how basic debayering interpolations can work, show examples of how adaptive interpolation can, statistically speaking, increase resolution beyond simple linear interpolation.. then people who are interested might gain actual understanding of the processes at play

The Alexa can record raw and non-raw .. the non-raw looks as good as it does, in part, becasue of the downsample from the spatially over-sampled bayer .. and the split signal gains for retaining low and high lights.. and maybe the log colour format.. all systems are complex , lets learn about them

 

[Graeme Nattress]

My main issue here is correct use of terminology, because there's been some serious abuses of this terminology in the past from people who should know better.

Graeme