"numerically honest 4k", Sony backhanded slap at Red

[Andrew M.]

My guess puts it around ~5700 x ~3000 if it were rotated 45deg back into normally oriented bayer pattern.

Actually if we know the sensor size it will be easy to calculate the equivalent bayer horizontal and vertical pixel count. Sure debayering process is very crucial but let's talk just photo-sites here.
I would trust that Graeme knows the best how to debayer by now.

It basically doesn't matter if pixels are placed at 45deg 20deg or 90 deg.

What does matter, is what are the dimensions of the sensor and how many square photo-sites are there. (we presume square pixels)

Anybody knows the dimensions of the active area or the whole area of the f65 sensor?

 

[Graeme Nattress]

For sake of argument, if you take the Sony's 4k resolution, they use twice as many pixels to achieve that in there rot45 bayer as with traditional bayer. So the factor to convert their "8k" to real bayer so we can figure maximal measured resolution from that number of pixels is to divide by root 2 = 5.65k. So what I'm saying is if their pixels were re-arranged to traditional bayer, they could have a 5.65k sensor, with no doubt a measured resolution and low aliasing figure of around 4.3k. Instead, I contend, that with their rot45 arrangement they'll never measure more than 4k and have higher levels of aliasing. But charts will tell us for sure. I see absolutely no way that sensor can measure any good percentage of 8k.

Graeme

 

[Andrew M.]

Also saying that f65 is 20MP and 10MP are green and than in the same sentence saying that there are 4096 x 2160 green photo-sites suggests that the active area has 16.87MP and the remaining 3.2 MP is look around area.

Andrew

 

[Deanan]

Also saying that f65 is 20MP and 10MP are green and than in the same sentence saying that there are 4096 x 2160 green photo-sites suggests that the active area has 16.87MP and the remaining 3.2 MP is look around area.

Andrew

They've stated in the sensor papers that the active area is 17.7MP although the FDtimes article said 18.3 but that seems to be a back calculation off the 4.2u pixel size which seems to be a rounded off number.

 

[AnthonyFlores]

For sake of argument, if you take the Sony's 4k resolution, they use twice as many pixels to achieve that in there rot45 bayer as with traditional bayer. So the factor to convert their "8k" to real bayer so we can figure maximal measured resolution from that number of pixels is to divide by root 2 = 5.65k. So what I'm saying is if their pixels were re-arranged to traditional bayer, they could have a 5.65k sensor, with no doubt a measured resolution and low aliasing figure of around 4.3k. Instead, I contend, that with their rot45 arrangement they'll never measure more than 4k and have higher levels of aliasing. But charts will tell us for sure. I see absolutely no way that sensor can measure any good percentage of 8k.

Graeme

Graeme, forgive me ... I'm not strong in this area ... so are you saying the NET measured resolution for the Sony after debayering is around 4.3k -- thus about equal to the Epic? I think you guys said in one of the recon threads that the Epic minimum would be 4.2k after debayer, and as much as 4.3 or 4.4k.

An "8k" sensor measuring roughly the same as a 5k sensor would be quite interesting indeed

 

[Graeme Nattress]

Anthony, I'm saying that if they arranged the same number of same sized pixels as a traditional rather than rot45 Bayer, they're net, with a decent OLPF and decent demosaic I'd expect around 4.3k. To me that is the best figure you can possibly get from that number of pixels and Green pixel to red/blue pixels ratio. However, they've gone with rot45 Bayer, and with that arrangement I expect 4k resolution (no more) and significant aliasing. Charts will tell us for sure though. Either Sony will release charts, and they've used Zone Plates in their theory demonstrations in the rot45, or I'll get a F65 and shoot charts.

Graeme

 

[Tom Lowe]

If aspect ratio is 1.9:1 and measured resolution is 4 - 4.3K, what would the measured resolution be cropped to 3:2 aspect ratio for IMAX?

How would this compare to the Epic? Epic is 16:9, no?

 

[Graeme Nattress]

Measured resolution would only be 4.3k if they used a real bayer though. It probably maxes out at 4k with their arrangement, and aliasing / resolution at that point will be OLPF dependent. I honestly can't see this camera measuring more resolution and less aliasing than an Epic, but I'm willing to be proven wrong.

Graeme

 

[Tom Lowe]

So in terms of using the Epic or F65 for 3:2 IMAX, it will probably be close, with a slight edge to Epic for resolution?

 

[Graeme Nattress]

I'd expect so, but charts will confirm this one way or another.

Graeme

 

[Andrew M.]

If aspect ratio is 1.9:1 and measured resolution is 4 - 4.3K, what would the measured resolution be cropped to 3:2 aspect ratio for IMAX?

How would this compare to the Epic? Epic is 16:9, no?

It will be 3.88K for 3:2 aspect ratio on f65.
This is all presuming 80% efficiency for the debayring process.

The sensor is 550 pixels narrower in comparison to EPIC.


But as Graeme mentioned, if they go with rot45 the debayering efficiency maybe as low as 65%

Andrew

 

[Tom Lowe]

If 1.9:1 is 4K, then how is 3:2 3.88K?

That doesn't add up.

 

[Andrew M.]

If 1.9:1 is 4K, then how is 3:2 3.88K?

That doesn't add up.

First we convert the 45 deg diagonal pixel count to horizontal.
We know that f65 has 50% of green photosites count of 4096 x 2160 and 50% of two other colors.
Without complicated math we can multiply this numbers by 1.5 to get the 90 deg conversion.

So it is 6144 x 3240 RAW pixel count sensor if we rotate the photosites pattern 45 deg CC.
Where 50% are Green 25% are Red and 25% are Blue.

The 6144 x 3240 sensor has 1.89 form factor.
The limiting vertical photosites RAW pixel count is 3240.
For 3:2 screen the horizontal pixel count will be than 3240/ 2 and time 3 = 4860
The usable active area of the sensor for 3:2 screen will be 4860 x 3240

Now with the 80% maximum debayer efficiency the measured horizontal resolution is:
4860 x .8 = 3.88K

80% is the best case scenario for bayer sensor so far.
As Graeme mentioned above rot45 pattern sensor may end up with 65% debayer efficiency.

EPIC is 5K RAW bayer pixel count camera, after debayering on the res charts we see good “honest”4K acceptable resolving power. This includes lenses, debayer process, OLPF, etc.

It is more logical explanation than mathematical and it comes from what I am reading here and from the experiments with RED ONE. Graeme can explain it better in details.

 

[Graeme Nattress]

Andrew, multiple by root2 not 1.5

Graeme

 

[Andrew M.]

Yes, it should be times 1.41.

Also I was thinking about 45 deg 90 deg and other factors.
Film doesn’t have ugly aliasing effect because particles are spaced on the surface of the film in random.
Good film has maybe in best 20MParticles.

By spacing pixels on the CMOS sensor in bit random, not 45 deg or 90 deg but semi random offset pattern of 20%-30% of photo-sites dimension we can eliminate aliasing.

Shifting the cells positions in semi random pattern and still maintaining good fill in factors will kill the aliasing once for all.
Don’t have to be really very random patter just odd-even 3 or 6 step pattern.

Used in telecom signalling to improve the S/N ratio.

http://www.ee.ucla.edu/~wesel/documents/3762.9076.pdf

This could be also achieved in software by semi random selection of adjacent photo-sites for calculating R or G or B pixel in the debayering process and dropping OLPF altogether.

http://en.wikipedia.org/wiki/Turbo_code



Andrew

 

[Graeme Nattress]

Well, film has a random grain pattern on each frame, which is pretty hard to do with a fixed sensor. You could emulate this with a higher resolution sensor, but at that point, you'd be lowering aliasing anyway by going to a higher resolution sensor.

But.... When treated well, as we tend to do, aliasing is not very much of an issue, and you get a nice high resolution. Aliasing is much more a problem with lower resolution systems and ones with poor OLPF choices.

At which point the added complexity outweighs the benefit.

Graeme

 

[Andrew M.]

In debayering of RGB pixel you don’t have to always pick the same neighbouring green pixel.
We have choice of 4 different green adjacent pixels that could be picked in semi random pattern.

But the best would be to shift photosites on the photomask by ½ of its dimension in random in all dirrections without letting fill in factor to go down.

 

[Andrew M.]

I asked the roof guy to put the slates in semi-random patter on the roof.
I think the sensor guys can move photosites like this on the photomask, can’t they?

No loss of the fill in factor as well and OLPF could be a history and a bit of film grain effect welcome in.

 

[Deanan]

I asked the roof guy to put the slates in semi-random patter on the roof.
I think the sensor guys can move photosites like this on the photomask, can’t they?

No loss of the fill in factor as well and OLPF could be a history.

You still have a fixed pattern, not a randomly varying one.

 

[Andrew M.]

This is just an example, the slates are in 3 different shades of gray.

When physically dimensioning the placement of the sensor photosites we have to disturb the horizontal line as well.
He did good job on randomizing the vertical line, in spite that all slates are 12” wide and 6” overlap.

Randomizing has to be picked carefully by rotating circuit area in relation to the photosensor area around the photosites.
Looking under this angle we can say that what OLPF is doing is introducing the noise to the regular pattern, randomizing the physical placement of the light on the sensor.

You can randomize both vertical and horizontal lines of photosites just by picking different start points for each line, what he did just for vertical lines.