Dragon 6k not eligible for UHD "Tier 1" camera?

[Noah Yuan-Vogel]

Let's use noise reduction as an example. AFAIK, all of the top CMOS Bayer CFA motion cameras do at least some NR early in the signal chain to improve signal to noise ratio. Out of all of them, RED does the least noise reduction in camera. Why? For some material, it may suit creative intent to have a noisier, "grainier", grittier texture. Moreover, less NR also means less attenuation of high frequency detail and greater fidelity to scene referred values. Good noise reduction techniques can really improve SNR, which is why most camera systems use NR more aggressively than RED. The downside is that once you toss data that the DSP decided was noise, it's gone. With an R3D file, you can apply NR with an eye to the particular material at hand, being more or less aggressive depending on the effect it has on the image in relationship to the intended look.

So you believe Sony and ARRI etc are denoising their RAW data and RED does not or does but to a lesser degree? Do you have any evidence or tests that support this statement? I have never seen any references to any of those manufacturers doing any noise reduction processing on their bayer data, but would be very interested to know if they do. Certainly compression tends to have a noise reducing effect, but if anything I'd think that would put ARRIRAW or lossless DNG RAW ahead of REDCODE or Sony RAW in terms of less signal processing.

 

[Milano Questa]

I love how Redusers automatically assume a test is flawed when another camera performs better than Red.

 

[Phil Holland]

I love how Redusers automatically assume a test is flawed when another camera performs better than Red.

I love how a somewhat flawed test can cause a domino effect towards an entire market that will or will not accept a specific camera which is in particular at odds with other entity's tests that accept a specific cameras.

Specifically that is my criticism regarding this particular article, interview, perspective, and quote from somebody who does impact a specific zone that does indeed film worldwide.

It's not exactly about being defensive in this case, it's more about broadcasting the correct information as there's been a little bit of something lost here along the way.

Again, it's more in criticism of what falls under Tier 1 cameras and what doesn't as there seems to be some confusion in there.

 

[Patrick Tresch]

I love how Redusers automatically assume a test is flawed when another camera performs better than Red.

Have you read the thread?

 

[Blair S. Paulsen]

So you believe Sony and ARRI etc are denoising their RAW data and RED does not or does but to a lesser degree? Do you have any evidence or tests that support this statement? I have never seen any references to any of those manufacturers doing any noise reduction processing on their bayer data, but would be very interested to know if they do. Certainly compression tends to have a noise reducing effect, but if anything I'd think that would put ARRIRAW or lossless DNG RAW ahead of REDCODE or Sony RAW in terms of less signal processing.

Noise reduction is but one example of the early stage processing that is common in modern cameras. The processing stage I am referring to is upstream from assigning discrete values based on voltages. Signal re-characterization prior to quantization is often employed to better utilize standard data storage methodologies. Every sensor is less than perfect, so every camera front end does standard compensation routines. The list goes on.

BTW, not all compression schemes reduce noise. An entropy codec like J2K or RedCode does attenuate high frequency detail, more so as the ratio increases, but not all codecs use the same techniques. MPEG2, for example, can accentuate false edges and create macro block artifacts that can make noise stand out - and don't get me started on banding.

Back to the original topic of this thread, do we know what compression ratio the tester's Epic was set to?

We can debate the intricacies of on board DSP in modern cameras if you like, but I'm more focused on why some tests show rather different results than others.

Cheers - #19

 

[Sabovic Adis]

There are "Authorized RED Dealers", therefore I suggest introducing "Authorized RED Resolution Testers"!
I nominate Graeme...:biggrin:

 

[Gavin Greenwalt]

So you believe Sony and ARRI etc are denoising their RAW data and RED does not or does but to a lesser degree?

Dig into the shadows of Arri footage and it goes "green". That's because they're trying to process out the noise. Arri definitely does some clever signal processing in the shadows to reduce chroma noise. Canon does notable noise reduction. Even in CR2 raw files you'll see pretty clear evidence of de-noise. For whatever reason, I never get Sony or BM footage so I don't know what they do. I assume they do something though. RED has been a little too-hands off IMO.

 

[Daniel Reed]

Well, if they are talking pre-debayer,
VVeapon 8K would qualify, but not Alexa 65!

In the above 6x6 bayer pattern example

R: 3x3
G1: 3x3
G2: 3x3
B: 3x3

Red 8K bayer pattern per channel resolution (8192x4320):

R: 4096x2160
G1: 4096x2160
G2: 4096x2160
B: 4096x2160


Alexa 65 "Open Gate" bayer pattern per channel resolution (6560x3100):

R: 3280x1550
G1: 3280x1550
G2: 3280x1550
B: 3280x1550

None would qualify... (Weapon 8k would though)

Hello,

I was suprised to read in the AFC that Dragon 6k sensor would be arbitrary not selected for "Tier1" UHD by the DPP (Digital Production Partnership).

Their argument is that it should have 3840x2160 pixels for each channel R/G/B. Now I don't understand why they should qualify Alexa65 that doesn't have 3840-2160 pixels for each channel R/G/B with it's 6560x3102 pixels?



Page 11.

Your math is off.....

A CFA with 8192x4320 (8k) has:

R: 4096x2160
G1: 4096x2160
G2: 4096x2160
B: 4096x2160

I think you divided by 4 but when considering linear dimensions you divide by 2.

Graeme

 

[Jacek Zakowicz]

Don't give them any ideas ;-)

 

[Hrvoje Simic]

As far as EBU is concerned, before starting bonfires and burning some folks, let's do some reading first:

3.1.4 Resolution
Camera resolution determines a camera’s Tier position. The measured video output resolutions of the various Tiers are as follows:
• UHD1 Tier 1 sensor resolution must be 3840 x 2160 in each of R & G & B. Where other sensor matrix technologies are employed, the actual resolution may have to be confirmed by measurement (EBU Tech 3335). As guide, a Bayer pattern sensor would need to be at least 5760 x 3240 to achieve full UHD1 resolution.

As far as BBC is concerned, the list states 4:4:4 as criteria as well.
In the context of broadcast drama that makes sense, so no 4:4:4 out is a factor which likely affects the decision of decision makers.

 

[Phil Holland]

Before starting bonfires and burning some folks, let's do some reading first:

Again, I have no issue with their Tier definitions. I have issue with a specific test that clearly missed a few things and an issue with the interviewed BBC fellow who is classifying cameras to specific Tiers that they themselves don't adhere to.

All of which has produced weird information which clashes against industry tests elsewhere.

 

[Patrick Tresch]

well, if they are talking pre-debayer... even 8K VVeapon wouldn't qualify (nothing on the market would).

8K bayer pattern per channel resolution (8192x4320):

You make a good point:

3840x2160 UHD Screens don't have 3840x2160 x 3 Resolution but 3840x2160 pixels x 3 variation of colors.

This is the big confusion they make. A colors on a Bayer pattern doesn't exist, but It does only after debayering. So the question is at what minimum resolution can the Dragon out resolve a 4k chart (chroma & luma) taking into account : the lens used, the olpf, the compression used, the debayer algorithm, the exported file, the resolution app that reads the data?
They speak of 5760 x 3240 pixels, I wonder how they find this pixels count?
As he stated, the tester didn't have a 4k chart.
Weapon max 16:9 resolution has 5617x3160 pixels. Was this argument used agains the Dragon?


Now he evaluated the F65 and the A65 ready for Tier1. I don't see the results in the link.


If we use their theorical 5760 x 3240 value, the CMO should have 18,66 MP.

Please correct me if i'm wrong.

Pat

 

[AndreeMarkefors]

To get the required minimum "3840x2160" res in the red/blue channels before debayer would require a 16KHD (15360x8640) sensor:

R: 3840x2160
G: 7680x4320
B: 3840x2160


Please correct me if i'm wrong.

The more likely scenario they are envisioning for broadcast—at least that's my impression—is a 3 chip solution. So, three discreet sensors of 3840x2160 each in the R, G and B.

That way you also maintain the UHD size internally with parallel processing and you won't have to design your processing specs to handle 16k…

 

[Patrick Tresch]

The more likely scenario they are envisioning for broadcast—at least that's my impression—is a 3 chip solution. So, three discreet sensors of 3840x2160 each in the R, G and B.

That way you also maintain the UHD size internally with parallel processing and you won't have to design your processing specs to handle 16k…

But a 3 chip solution is optically not a better option.

 

[AndreeMarkefors]

But a 3 chip solution is optically not a better option.

The only thing I'm saying, after having read through a couple of the EBU reports, is that the specs are not necessary in relation to any existing camera or technology, but rather just list a few technical specs that you need to reach a certain tier.

At some point when the describe the full resolution channels, it's in conjunction with 3 chip cameras.

What I would call 'full oversampling'—as Daniel describes above with the Bayer pattern and the 2x resolution needed—quickly becomes cumbersome at the forefront of the next higher resolution jump: SD-->HD-->UHD-->8k UHD.

Note, I'm also not saying that full resolutions channels are needed. I'm just offering another solution that isn't single chip Bayer—that seems easier to achieve.

 

[Sabovic Adis]

Aren't they confusing resolution with sharpness?
Resolution => pixels
Sharpness => linepairs
Right?

 

[Graeme Nattress]

Your math is off.....

A CFA with 8192x4320 (8k) has:

R: 4096x2160
G1: 4096x2160
G2: 4096x2160
B: 4096x2160

I think you divided by 4 but when considering linear dimensions you divide by 2.

Graeme

 

[Graeme Nattress]

Resolution and sharpness are indeed not the same thing. You could record a VHS tape out to a 4k file and have 4k resolution, but only VHS sharpness

When we're talking the resolution of a chip or a file format, we're talking pixels and we can be very specific and say the resolution of this bayer CFA sensor is x by y pixels. When talking sharpness, that's a measured quality, and thus we must measure it..... Standard is a linear sine zone plate, from which we can read MTF. We use the zone plate because it also shows aliasing. Our perception of sharpness is proxied by MTF50, the point at which our contrast has dropped to 50% of which it started with. The higher the MTF50, the higher the perceived sharpness. For limiting resolution we can use MTF10.

Graeme

 

[Sabovic Adis]

Thank You, Graeme, my point exactly! :thumbsup:
So, what resolution are we (they) talking about: that of a sensor, or that of a debaeyred RGB file? :biggrin:

 

[Graeme Nattress]

From my understanding they're talking measured resolution of the resulting image.

Graeme