How does an F3 output an "RGB 4:4:4" data?

[Steve Mullen]

Likely a dumb question, but -- the F3 is a single chip camera which I assume uses a Bayer filter. IF it has only 3.4MP -- and not Canon's 8MP design -- then as far as I know, there is no way to get a 4:4:4 signal because each pixel does not have all 3 colors.

So how does an F3 output an "RGB 4:4:4" data?

Which raises another question.

I have assumed the RGB data comes from the sensor. Which if true would make the output sRGB not RGB -- like RAW from a still camera. But, if the RGB data is not from the sensor -- then where does it come from?

Note: I assume that 16-bit sRGB from the sensor is deBayered fairly immediately to 4:2:2 10-bit YCrCb and all DSP (image adjustments) are made with YCrCb. If I'm correct, then either YCrCb is converted to RGB before being output via dual HD-SDI or it is tapped before being converted to YCrCb. In the former case, image adjustment would be reflected in the RGB output. In the latter case, image adjustments would not be reflected.

Reading the F3 manual I can't find anything about image adjustments when the RGB firmware is installed.

[Chris Medico]

The only adjustment available when in RGB mode is white balance. 3200 or 5600 presets. Thats it.

[Steve Mullen]

Is WB done by an optical filter or electronic adjustment?

If electronic, then either gain is adjusted for each photocite as its signal comes from the sensor or the sensor is deBayerd to RGB and each channel's gain adjusted.

In any case, it means the RGB is output before the conversion to YCrCb.

So it seems either sRGB or RGB is output. So is there one serial digital signal -- R then G then B -- and the deBayering is done in software (like an RED) or the deBayer is done in the camera and RGB is output.

My hunch is the de-Bayer is done in the camera -- so the F3 does not output RAW. But, then how can Sony get 4:4:4 from a deBayered sensor?

PS: The same question is relevant to an ALEXA. It can write ProRes 4:4:4. How can 4:4:4 come from a deBayered sensor?

Perhaps, "4:4:4" simply means "equal" R and G an B "detail" after deBayering. That is -- each de-Bayered pixel has an RGB value.

If I'm correct, then this kind RGB data has less real resolution than would a 3-chip camera with the same resolution chips.

[Chris Medico]

It is electronic. The only movable filter in front of the sensor is ND.

[David C. Williams]

Off the top of my head I can't think of a single video sensor type that produces a full RGB pixel from one photo sensor? I could be wrong (not for the first time), but they all collate and interpolate from at least three or more photo sensors. Even your eyes do something similar.

With that said, arguing that this or that sensor can't do true RGB 4:4:4 is kind of pointless as none do strictly. They all do it in different ways, some better than others. All it means in this context is that each individual pixel in a cameras output can be unique compared to the one next to it. 4:2:2 or 4:2:0 just means a reduced colour resolution with some colours derived from adjacent pixels.

The only sRGB I'm aware of is a colour space standard created my HP and MS. Nothing to do with what comes out of a sensor directly.

[Steve Mullen]

I agree.

With a 3 chip camera, each pixel has a sample from 3 photosites. With 2MP sensors there are 2M R and 2M G and 2M B samples. Each of the "2M pixels" has a R, G, and B sample.

After converted to YCrCb, 4:4:4 is possible. Normally, however, every other Cr and Cb value is discarded yielding 4:2:2.

With a 1 chip camera, each pixel has a sample from 1 photosite. With a 2MP sensor there are 1/2M R and 1M G and 1/ 2M B samples. We both agree these "2M pixels" cannot have have the same resolution per color as the "2M pixels" from a 3-chip camera. So after de-Bayering, although there will be 2M YCrCb pixels, resolution per color is much lower. So, although 4:4:4 RGB output can come from these 2M pixels, the image's resolution is lower than the 2M 4:4:4 from a 3-chip camera.

As I see it, by starting with a 3.4MP sensor -- after deBayering the resulting 2M has the resolution as the 2M 4:4:4 output from 3-chip camera. (The math works.) Thus, the 4:4:4 does have the same resolution as from a 3-chip camera.

This also means the 4:4:4 RGB output is obtained after deBayering. Then the conversion to YCrCb happens.

Does this all make sense?

[David Heath]

First off, the F3 does not - can not - output "true" 1080 4:4:4. IF, by that, it is meant R,G,B each at 1920x1080 resolution.

I like to think of the colour space as a series of "containers", and let's (very simplistically) think of 4:4:4 as "2 litre" and 4:2:2 as "1 litre". Now let's go further, and think of the output signal as "the liquid", and let's assume the F3 is capable of producing 1.5 litres of "liquid".

What do you do? You must either use a 2 litre container to preserve the lot (inefficient in terms of storage space) or use a 1 litre container and discard a third of the liquid (efficient in terms of storage, but you're not taking full advantage of the sensor).

Coming back to images, the F3 sensor seems to be about 2456x1372 photosites effective. After deBayering you can expect a luminance signal of about 80% that (say about 1965x1100, nicely matching 1080 specs) and chrominance of about 50% the sensor specs (say about 1230x690). That chrominance data more than fills the 4:2:2 "bucket" (at least horizontally) - but leaves the 4:4:4 one partly empty. And unfortunately, there is no size of "container" between the two! :-)

Key is to bear in mind that each and every of the photosites can contribute luminance information for a black and white image, but none singly can contribute unique chrominance. Hence why deBayering gives better luminance than chrominance resolution - and since this is how the human eye also perceives, normally it's no problem.

[Steve Kalle]

Another 'smart' Steve around here ;)

I, too, have been critical of Sony's claims to the 444 output with so few pixels, which is why I have been perfectly happy recording its S-Log 422 output with 14 stops of DR at ISO 800 (people need to realize that every digital camera -photo & video- lose DR as the sensor's gain is increased. For example, take a look at DxOMark's DR tests of DSLRs at each ISO. Some cameras can go from 12-13 stops of DR at ISO 100 down to only 9 stops at ISO 3200. This is one of the reasons why I don't see much use for 'hi' ISO on a camera like the C300 with its alleged 20,000 ISO (with only 9 stops in normal gamma and 12 in log at its base 850 ISO, there is very little DR left at such a hi ISO). However, this doesn't apply to high-end cinema cameras such as the F35 and Alexa which retain the total amount of DR at each ISO but the amount of DR below and above mid-grey vary - eg., from Arri's website "EI 160 (+5.0 / -9.0), EI 200 (+5.3 / -8.7), EI 400 (+6.3 / -7.7), EI 800 (+7.4 / -6.6), EI 1600 (+8.4 / -5.6), EI3200 (+9.4 / -4.6)"

After almost of a year of the 444 upgrade being available, I have yet to see a test demonstrating an advantage and/or difference between S-Log 444 and S-Log 422.

About the Alexa - it records to Pro Res at 1080p (10bits) which means it records a debayered/down sampled image whereas its ARRIRAW records 2880 x 1620 (at 12bits).

[David Heath]

Originally Posted by Steve Kalle

I, too, have been critical of Sony's claims to the 444 output with so few pixels, ....

The claims are valid in so far as it does record a 4:4:4 CONTAINER, it's just that it doesn't contain 1920x1080 each for R,G & B resolutions - which is what is widely ASSUMED when you see that magic "4:4:4".

It's nothing new - many cameras claim "1080 recording". It's a true claim in so far as they record a 1080 signal - but away from the higher end, in many, many cases they struggle to manage anything equating to a 1 megapixel image, not the 2 megapixel that "1080" implies.

444 recording on the F3 should give a noticeable improvement over 422 on horizontal chrominance resolution - but don't expect it to be anything like the doubling that the 444v422 would lead you to expect.

[Duke Marsh]

But the fact remains that if you're pixel peeping on an FX shot the F3's 4.4.4 image files do look different than the 4.2.2 images.

Just not enough to pay much attention to in a regular shot. Especially with S-Log.

[Andy Shipsides]

When they say RGB 4:4:4 the are strictly referring to the video output of the camera. The dual link output produces an RGB signal, regardless of how it was produced it is indeed RGB. By the way all RGB images are free of chroma subsampling, so calling it 4:4:4 is a bit redundant. We could argue that many, if not post, cameras don't actually capture equal quantities of Red, Green, and Blue pixels. However the 4:4:4 / 4:2:2 description is talking about traditional chroma subsampling in a video signal or compression.

As Steve said, 4:4:4 give you better chroma sampling overall, but not a huge improvement in the image otherwise.

Andy

[Alister Chapman]

RGB tends to give better performance than YCbCr with highly saturated, high brightness colours and also there is less cross talk between colours, so this tends to make it easier to grade. However as the source data in the F3 comes from a Bayer sensor this advantage is very small. When I do shoot RGB on the Gemini it does have a slight edge over the ProResHQ YCbCr from the Samurai. You can't really see it first generation, but after any heavy grading the RGB just looks a tiny bit nicer.

I'm happy to work either way. Smaller files when I'm out in the field with the YCbCr Samurai, but when I want ultimate quality, I can use the RGB Gemini.

[David Heath]

Originally Posted by Andy Shipsides

By the way all RGB images are free of chroma subsampling, ....

I suppose it depends how you define "chroma subsampling". It's true that RGB recording doesn't involve subsampling *at the recording stage* (which is what normally gets referred to) - but think how the image is actually formed.

It starts off as a Bayer matrix of about 2456x1372, which then gets deBayered and downscaled. Even if the process never involves the use of component signals, it still remains true that the image has higher resolution for luminance than chrominance - even if it is carried as R,G,B signals, not component. Put another way, "true 4:4:4" recording has the capacity to capture full 1920x1080 resolution signals for each of red, green and blue. But the sensor doesn't have the ability to produce that. If you use the term "chroma subsampling" to mean "colour resolution information less than luminance", then that is what the sensor gives.

We could argue that many, if not post, cameras don't actually capture equal quantities of Red, Green, and Blue pixels. However the 4:4:4 / 4:2:2 description is talking about traditional chroma subsampling in a video signal or compression.

And ironically the C300 sensor IS capable of giving "true 1080 4:4:4" (be it RGB or component) - but the camera does not have the means to record or even output it.

In practice, we're all saying roughly the same thing. The RGB recording system will give a performance improvement over component 422 - but not as big as may be expected from a simple look at the specs.

If the C300 had the 4:4:4 RGB option, then I'd expect a much greater difference between that and 422 than with the F3.

[Alister Chapman]

Playing with my C300 and F3 and I'm really struggling to see a visual difference between the 50Mb/s 4:2:2 from the C300 and the 35Mb/s 4:2:0 of the F3. Not done any grading tests yet to see how the C300 holds up compared to the F3. What will be interesting is to see how the Canon C-Log at 8 bit 4:2:2 holds up compared to 10 bit 4:4:4 from the F3.

Canon C300 and Sony F3 footage to download. | XDCAM-USER.COM

[Doug Jensen]

Alister, that does not surprise me since it is virtually impossible to see a visual difference between the F3 recorded at 35Mb/s 4:2:0 to SxS cards and the F3 recorded at 50Mb/s 4:2:2 (or even 100Mb/s) to a NanoFlash.