Now that Red is the acknowledged Future King...

[Lauri Kettunen]

It sounds like Bruce says the well knwon fact that one can employ numbers (statistics) to emphasize the message one wants to make. A simple example reveals the point. One can put it this way to emphasize the message in another way:

Canon XL H1, three 1440x1080 sensors:
G channel: 1.5 M pixels
R channel: 1.5 M pixels
B channel: 1.5 M pixels
In total 4.5M pixels

2K camera with 2048x1152 CMOS sensor:
G: 1.18M pixels,
R: 0.59M pixels,
B: 0.59M pixels.
In total 2.36M pixels.

4K RED with 4520x2540 sensor:
G: 5.74M pixels
R: 2.87M pixels
B: 2.87M pixels
In total 11.48M pixels

Now, compare if one puts it this way:

XL H1: 1.5M pixels
2K: 2.36M pixels
4K: 11.48M pixels

Bruce does have a point, doesn't he; In the rigourous sense of the word, in fact, the deBayer process is about uprezzing, isn't it? That is, one creates -the technical term is by "cultivated guessing"- data which is not actually captured by the sensor. This means, if you know the algorithm, you can always deliberately create an image in which the algortihm will somewhat fail. From the practical point of view this is water on the duck's back, for the odds to meet such situation in practice is negligible. So, one may as well say this is all hair splitting.

 

[Graeme Nattress]

I think of the bayer pattern on the sensor as a clever form of 3:1 compression. The bayer demosaic algorithm being decompression.

It's possible to feed any compression a "pathalogical" image.

However, one of the hardest images for any camera is a zone plate, (they show up resolution and aliassing quite clearly) and that image didn't confuse the demosaic algorithm.

Graeme

 

[Stuart English]

Resolution

Resolution

Last comment on this.

You might say that statistics can prove anything, but the fact is that through appropriate image processing each pixel in the Mysterium sensor can be used to generate both a luminance signal and a color signal. So your number of derived samples (and hence resolution) is based on 11.4MP.

That is not true of a 3 x 2.1M sensor. It that case, its true that each pixel on each sensor can also be used to generate a luminance signal and a color signal but as all the CCD's are co-sited, the total number or derived samples (and hence resolution) will still be based on 2.1MP, not 6.3MP.

 

[Bruce Allen]

Bruce, what are you babeling about ????
On another note, How are the monitors coming along ?
Aloha
-A

Poi Boy, okay, I am shutting up. I have moved on to post crazy photos of my rig plus monitor.
(http://www.reduser.net/forum/showthread.php?t=3918)

Graeme, I agree about Bayer being smart 3:1 compression. But then 3:1:1 is also a pretty smart move and you guys rip them apart for using that in your paragraph I don't like

Now I'm going to leave this argument and next time someone I refer to your site asks me about your paragraph (and they will, because most of your prospective customers are smart and inquisitive and are also very jaded after years of Sony marketing) I can tell them I have posted on the forums and you guys responded.

Bruce Allen
www.boacinema.com

 

[Lauri Kettunen]

As RED is a fantastic camera this will also be my last comment as well.

This twist of words is hidden here:

derived samples

The bayer demosaic algorithm being decompression.

The total number of R, G, and B samples is 11.4M, and all other data is, in fact, lost. And formally speaking, there is no way to generate it back afterwards. However, if one makes a reasonable assumption, say that the colors change from one pixel to another in somewhat regular manner, then you can create based on the assumption employed the lost/missing 22.8M RGB pixel data. That's the deBayer process.

Formally, in the technical sense and in words of logic, the sensor creates a so called equivalence class, meaning there is a large number of images corresponding to the one and the same RAW-data of the sensor. Now, there is no way to pick the canonical 3 x 11.4M image RGB-data from the (equivalence class of) 11.4 samples one has available. The rest of data is lost and lost for ever. But now, there room for cultivated guesses one can make here, such as the lost data was somewhat regular. Based on this the deBayer process constructs a so called representative of the equivalence class and the representative -or you may call it "proxy"- is close enough to what the canonical data would be.

So, my answer to Graeme is, if you consider the sensor sampling as compression, it's definitely a lossy compression. The answer to Stuart is, yes, they are indeed derived samples but not genuine or real samples. For this reason one may well call the deBayerian uprezzing It's just a matter of taste which words one employes here.

All this is indeed just hair splitting, if one is not able to come up with a practical case where all this had a meaning. And I think this is basically what Stuart and Graeme say: those cases where the reconstruction did not work are "pathological". However, there is a chance that the issue is not studied in depth for typical resolution charts fall into the category of "somewhat regular images". But, if the resolution charts were fractals, such as the snow flake (made of an equilateral triangle, having an equilateral triangle on each side and so on) then the basic assumption of somewhat regular data did not hold any longer.

If -and I emphasize, I have no idea how you do it- the RED deBayer process relies on the assumption of "somewhat regular data", then RED and a 3-sensor 2K camera shots on fractal resolution charts should show the 3-sensor approach has its advantages.

 

[Graeme Nattress]

The data is not totally lost though.... Because of the anti-alias filter. If you have a camera without an anti-alias filter you get nasty artifacts. Because of the necessary anti-alias filter, a 3 chip system does not have it's rated pixel resolution either. Some cameras omit or don't have adequate filtering and produce artifacts. The bayer process is not an uprez as the output resolution is not greater than the input resolution - it's a decoding or reconstruction process.

Another way to look at things is "efficiency". For us to make a 12mp image, we need 12MB * 12 / 8 = 18MB. Say the measured horizontal resolution is 3400 lines (not unreasonable). To achieve such a measured resolution on a 3 chip system you'd need 6.5mp sensors, giving a total of 6.5MB * 3 * 12 / 8 = 29.25MB. In that respect our "compression" ratio of bayer to 3 chip equivalent is 1.6:1 not 3:1. And of course, the measured resolution of a 3 chip 6.5mp system would be lower than that which I've assumed as it too would need and optical low pass filter. To get the resolution the same as our 12mp sensor, it would probably need more like 8mp or 10mp to account for the filtering.

Graeme

 

[James T Mather]

It seems to me that some people have too much time on their hands.

 

[Desert Rune]

After mulling over what Bruce has to say, I'll agree with him. Red has shown the goods. There's no more need to "prove" to the skeptics of what Red is, and so if the product introduction page goes through a rewrite, it would benefit us all.

All this tech talk about Bayer, sampling etc. doesn't really matter as long as we know the technology produces stunning imaging.

 

[Wayne Morellini]

I think of the bayer pattern on the sensor as a clever form of 3:1 compression. The bayer demosaic algorithm being decompression.

Graeme

There are certain domains in the data that overlap making debayering more realistic, but there is something even more aggressive I was thinking about doing years ago, before Red even started I think. 1/3rd three color pixel shift, needing three times less pixels again. So a three chip 2k chip can render 18mp, a nice amount. It is a shame that the three layer sensor manufactures have dropped the ball in terms of high quality video. Sigma sees it as unimportant, but if I buy an expensive still camera camera I want high quality video, and preferably 4k video. The Panasonic gh2 is one of the few that cones close, and that is after the hack, and it is cheap. What I could do with a competent ultra low crosstalk, high snr and dynamic range, 50 to 100fps, 100% fillfactor, bs 4k 3 layer sensor (closer to 100% the better). The patents must be running out, You could still use this for red code Bayer, by dropping the other channels, but I same up with some other compression techniques, at the least you can store the difference from the assumed difference between colored pixels. In this you assume the value of missing pixels of a color as an average from surrounding pixels of that color in Bayer or other domain (I do have more advanced techniques but am keeping them to myself as I do not remember if I revealed them or not) when the result from the sensor is different from the assumed, you store the difference as a general combined values of colors, or as individual channels alongside the Bayer to improve accuracy. There is a lot of extra stuff I explored. I was thinking of a Bayer sensor design that registered the primary and left over light per pixel.

 

[Wayne Morellini]

Forgot, what professionals don't tell us, is that the more colors you use in combination, the more accuracy you can get, and the more pixels you can get out of shift. In real life this might not matter too much in recording, but it has uses (very important) and at some point the quality of pixels from shift is going to break too much. But I wonder what we could get out of complimentary layers, or at least infrared and uv for security scientific study, which I wanted to do. However, recently I same up with a method to record virtually infinite seperate frequencies, but I can't remember how, I had another idea for this years ago too. Some of the stuff I have been theorising would be hard to shake a stick at, there is so much of it and years and years beyond what is happening. My scientific research is more radical. That's how you know a real businessman, how quickly he jumps on that to explore making money. There are not too many real businessmen out there.

Some wisdom there, between the lines.

 

[Gavin Greenwalt]

That line is still a pet peeve of mine as well. Pixel shifting and Bayer patterns seem like at most 1st cousins of the exact same system. They both use spacial offsets of different channels to interpolate the other colors. Which is in of itself a form of uprezzing.

"Pixel shifting not spoken here! (We use pixel sheefting.)"

 

[Meryem Ersoz]

just get rid of it and state the obvious - "most frickin beautiful images available on the planet today at this or any price point"

done and out.

 

[Graeme Nattress]

Zombie threads!!!!!!!!!!!!!!!!!

Gavin, the situation was, way back then, that things like pixel-shifting were done without informing the consumer that they were going on. Some companies were very cagey about the sensor resolution and how they achieved their output image. The difference being, we've always been very open about how we use a Bayer Pattern CFA sensor, what the pixel dimensions are and what the measured resolution you can expect is.

Graeme

 

[Wayne Morellini]

Whoops, I didn't realise this one was an older thread too, going have to apologise to Evin.

Threads that are too old should be archived.

 

[Matt Ryan]

lol I think Wayne is on a roll today hahaha

:auto:

 

[Wayne Morellini]

Ha, ha, funny thing is that is the color of my Car, and it is a Turbo diesel, and I don't drive it like that (not saying I could not).

 

[Jeff Kilgroe]

Threads that are too old should be archived.

I have suggested that to our admin gods on more than a few occasions in the past. Most the time threads get resurrected, it's accidental. Usually brought on by a quick reply after a search and bam, there it is. Often times, it's not such a bad thing in terms of a post or two that are relevant to the latest addition to the thread... Unfortunately, so many of these threads that come back to life contain old or outdated info. Too many occasions where we actually see a 3 year old thread pop up with old specs or info on something and the next thing you know, people are quoting it elsewhere on the forum or it's causing confusion because they don't look at the dates.

...If I had my way, any thread on these forums that goes for 90 days without posting activity will automatically lock itself. That's a rough amount of time off the top of my head, but probably not too far off the mark. It's usually better if someone starts a thread and then links to the old thread, or better yet, specific posts, for the necessary references.

 

[Bruce Allen]

Often times, it's not such a bad thing in terms of a post or two that are relevant to the latest addition to the thread... Unfortunately, so many of these threads that come back to life contain old or outdated info. Too many occasions where we actually see a 3 year old thread pop up with old specs or info on something and the next thing you know, people are quoting it elsewhere on the forum or it's causing confusion because they don't look at the dates.

Totally true. Plus they also confuse the guy who started the thread 4 1/2 years ago, haha!

I was like "WTF, I don't remember writing this! Did I sleepwalk off to my computer and go and post on REDUSER last night?"

I love the camera, love the people, love the footage, love the community.

At least that part still holds true And the Future King bit!

OK - maybe 2007 Bruce knew what he was talking about

Wow. 2007. I think that was a few months before we did the ol' Ringo video. Crazy!

Bruce Allen
www.boacinema.com

 

[Jay Birch]

quite funny to start a thread about hyperbole with the title "Now that Red is the acknowledged Future King..."

 

[Emanuel A.]

Yes, Bruce is a nice and interesting guy but I concur that to see him to spend his energy in this thing is a waste of time ;-) His talent and enthusiasm deserve much more (E