How can a single-chip camera claim 4:4:4 color?

M

Mark Sullivan

Member
How can a single-chip camera claim 4:4:4 color?

With a Bayer-pattern filter, it takes four pixels to get complete color information, right? So why is that any different from the chroma sub-sampling in normal Mini-DV compression? Why isn't 4:1:1 the absolute best that any single-chip camera can do, at least at full resolution?

I'm willing to belive there's something I don't understand about this, or there would be no point in using 3-chip sensor blocks for Mini-DV cameras since the output is going to be 4:1:1 anyway.

Can anyone help me with this apparent contradiction?
 
I'm not too up on the technical, but I believe it's the difference between a CCD sensor and a CMOS sensor.
 
That doesn't make any sense. No mattter how the underlying photodectors are made, a single-chip camera does not get full color information from a single pixel. It either gets R, G, or B and an interpolation algorithm combines these with the signal from adjacent pixels to interpolate the color.

My point (of confusion) is that the output of this interpolation comes up with a color value for each pixel, but wouldn't the overall bandwidth of the color information still be 1/4 of the total image bandwidth?
 
I hope we're not claiming 4:4:4 at full sensor resolution. That would not be the case for us or any other single chip bayer pattern sensor camera.

In some modes we deliver full 4:4:4 RGB though - "2x over-sampled HD in 4:4:4" because we're 2x oversampling, using 4 pixels to make up either 2k or HD. We can do this because we have many more pixels than we need for most resolutions.

Now lets look at full resolution or cropped modes, say 4k. At that resolution, it needs to be bayer de-mosaiced to be visible as RGB. There are many ways to de-mosaic, but good ones tend to give > 70% of the linear resolution out as real resolution, so 4k 4096x2304 gives out > than 2867x1612 real full resolution, although it will appear in an image of the full resolution of 4096x2304. It's the difference between measured resolution and pixel resolution. You can dub DV to HD and you will get a measured pixel resolution of 1920x1080, but it still has DV resolution, if you get my analogy. Now this is the same for all of use single sensor cameras, and Arri do a very good job of explaining it on their website. A bayer pattern image of 1920x1080 leads to about full 720p resolution, etc.

Also, the optical filtering in camera does some degree of spreading the colours over the pixels of the CFA, which doesn't happen in the decimation of 4:4:4 to 4:2:0 in a DV camera's processing.

Now, on a large chip, at high resolutions, bayer pattern will look better than you'd think from the above explanation, as lenses and reality limit the resolution coming into the sensor. You just have to look at the images from a top end Canon or Nikon DSLR to see that although bayer sounds bad, in reality, with good algorithms, it isn't.

Now onto DV....

NTSC DV is 4:1:1 but that's not a good analogue to bayer pattern. http://www.lafcpug.org/Tutorials/basic_chroma_sample.html

However, PAL DV 4:2:0 is. So for NTSC yes, you'd need 3 chips because of the pattern of the chroma sampling. But 3 chips at these low resolutions give advantages - some oversampling and they tend to be bigger chips. 1 chips cameras tend to have very small chips. As the resolution gets higher, any issues with bayer pattern rapidly diminish, and having one, nice, big chip becomes tremendously advantageous. Or another way to look at it, is the little Sony 1 chip HDV camera has the same resolution as it's big 3 chip brother, and some think it looks better too.

I hope that helps you,

Graeme
 
Graeme_Nattress said:
I hope we're not claiming 4:4:4 at full sensor resolution.
Click to expand...
I think the original "teaser" website certainly implied that ("2540p. 4:4:4." - though they never explicitly said they'd do them together...), but the wording in the new one does serve to clarify that notion. I think even 2K at 4:4:4 is amazing, and I doubt many of the "independent" crowd is going to be shooting 4K for awhile anyway, so it's not that big of a deal. The format options page (http://www.red.com/product/format-options/) still doesn't list 4:4:4 as an option for 1080, however, and I think there is just a little general confusion around because of some of these inconsistencies. That being said, I don't believe for a moment that the people behind this project are going to allow the camera to do anything short of amazing things, and I'm confident that the priority of image quality is (and has always been!) #1. Keep on truckin!
 
Last edited:
At 4k it's not 4:4:4 or any x:x:x. It's raw bayer pattern and it is what it is. However, yes, at 2k scaled it's 4:4:4 in every meaning of the notation and oversampled too.

Remember, we're in development and things can and will change. I hope this explains the issue though.

Graeme
 
Thanks for the 70% figure, Graeme. I've been wondering what the approximate "real" (de-Bayered) resolution is for a while.
 
I think I got that 70% from a BBC R&D paper, but the nice thing in reality is that it actually looks better than that! And remember, > 70%, so that number can be higher although it can never reach 100%, and it all depends upon the algorithm used. Some produce higher resolutions, but have higher noise or deal with colours worse, and some produce softer images with less artifacts or none at all.

Graeme
 
Last edited:
This discussion clears things up tremendously.

I remember when Nikon added the optical low pass filter in front of the 2MP sensor in the D1. It made an incredible difference. I did lots of testing on that camera. I'm excited to learn the Red has such a filter.
 
Every camera needs an OLPF. It's vital. Nikon usually make their's too weak, and get some extra resolution at the expense of aliassing. Canon usually get their's just right. I've never seen any nasty aliassing from them. Some high end medium format Digital cameras have no OLPF at all, and they alias a lot, especially on some architectural or clothing / fine weave shots.
 
I think im going to make this a sticky... thanks again Graeme for shedding some light, honestly and openly. You should be commended.

It must be great to be in a situation where you have more pixels then you need rather than not enough pixels. Having not enough pixels to hit a resolution is where you start to get in trouble, as you not only need to interpolate actual pixel resolution, but the color values within those artificial pixels.
 
The question of how much "actual" resolution you get from a Bayer mask came up in a thread before. At the time I posted that it was 70% (ish), but remember that someone who was obviously more involved with the technology than I am said in real terms they were seeing nearer 90%. Not sure who it was though.
 
Coincidentally, yesterday I asked Dalsa about maximum RBG output resolution from the Origin. I received this response: "We do our highest quality bayer reconstruction to 4096 x 2048 (4046 x 2048 if you cut out the 50 black pixels) at 16bits per RGB pixel."

So Dalsa can achieve 100% of sensor resolution from it's reconstruction algorithms. They use a single CCD with Bayer striping.

I was surprised. So much for generalizations.
 
I think everyone obtains 100% of sensor resolution when they demosaic the output of a Bayer-pattern sensor. This just means they get back the same number of pixels that they put in. The point of the number: 70%, 90%, etcetera is how good the reconstruction is compared to a real 3-color per pixel sensor (like a 3-chip sensor block).

I'm sure the success of the demosaic has a lot to do with the image. This may be one reason the figure of merit varies so much depending on whom you ask. There's probably not one perfect algorithm for all images. But there are also not a lot of images where the color bandwidth in the image truly needs to be as high as the luminance bandwidth.

The one application where you need all the color bandwidth you can get is chroma-key. I could see keying and compositing using the 4K raw output of the Red One, even though I would eventually scale down to 2K for the rest of the workflow.
 
Mark, that's a rather off-hand dismissal, because DALSA is the only company I know of that claims to be capable of projecting true 4K RGB images from a 4K sensor (apparently it takes 4 x 4:4:4 HDSDI connections for data transfer). As Bayer striped sensors appear to be the way of the immediate future, we all need to understand their capabilities better.
 
Last edited:
16bit at 100% resolution - I'll believe that when I see it. That would imply a >96db SNR (given their chips aroun in the >70db SNR stats according to the Dalsa web site), which I don't quite believe, and 100% RGB reconstruction seems to imply magic.
 
Last edited:
Graeme, rather than dismissing it as impossible, perhaps it would be useful for you to contact DALSA directly to get the facts yourself because it's ultimately an important issue and has relevance to RED. (If DALSA can do it, perhaps, RED can too?) I wanted to understand more about the various implementations of Bayer filtering and reconstruction, and especially about the claims of "real 4K" resolution after reconstruction because I too believed it wasn't possible, but that was DALSA's reply. I'm always eager to be enlightened by brighter minds, including yours, but facts are the best way to do that. If the info I received isn't convincing, then by all means make your own enquiries of DALSA. I suspect the board is interested in what you find out. I know I am.
 
Well, I've read just about every scholarly paper on de-mosaicing there is, and nobody can get 100% RGB equivalent resolution out. Given that Arri correctly, IMHO, compare RGB to bayer resolution, I'd have to say that perhaps Dalsa didn't understand your question to mean quite what we think it means. And if Dalsa are doing something not covered in the vast knowledge of what bayer can do, they're hardly likely to tell me what it is they're doing...

Or you can read Dalsa's literature that says : "Compared to a monochrome
sensor with the same pixel count and dimensions, the mosaic
filter approach lowers the spatial resolution available by roughly
30%, and it requires interpolation calculations to reconstruct the
color values for each pixel."

http://www.dalsa.com/dc/documents/I..._Whitepaper_Digital_Cinema_00218-00_03-70.pdf
 
You must log in or register to reply here.
Back
Top