sony

[Eric Lange]

Eric, don’t get excited. They have to differentiate themselves somehow and when technology is pushed to the limits like this sometimes it is difficult to differentiate.

Guess what, I wouldn’t mind to have two boxes of Shreddies on the shelf in the kitchen.

Imagine, you wake up and you look on the shelf with two boxes of Shreddies side by side one square and one diamond.

Which one would you pick if you eat squares Shreddies for past 5 days?

That’s a very funny but interesting point.

It not that I’m overly excited apart from trying to give up smoking and drinking way too much coffee…

I am worried that I am going to be seduced by globally shuttered wicked fast low SNR diamond Shreddies!

 

[Graeme Nattress]

My layman's non-technical perspective would say that pointed at a circular zone plate, a Bayer-sensor turned 45 degrees shouldn't show any change in aliasing, but in real-world shooting, there may be a slight advantage since a number of grid patterns encountered (like shooting TV sets) tend to be vertical and horizontal, less often diagonal. But I'd want to test that theory. And it wouldn't help with a grid pattern in a shirt, for example, where the collar and sleeves cause the pattern to turn diagonally.

I don't think so, as if you work through the math, you get an equivalent traditional Bayer array of 5792x3258, which is more than enough to get a 4.5k measured resolution with negligible aliasing.

Graeme

 

[Eric Lange]

I don't think so, as if you work through the math, you get an equivalent traditional Bayer array of 5792x3258, which is more than enough to get a 4.5k measured resolution with negligible aliasing.

Graeme

The different targets will measure the same value for resolved resolution. On a circular target on a small part of the tangent of one of the circles you will see a small “break up” that’s not a big deal or very noticeable. However if you have high frequency long horizontal features i.e. long thin horizontal lines and the CCD array is within a half a degree of rotation of that horizontal line, then as you move the camera slowly (particularly a tilt) you get flashing aliasing artifacts along these imaged horizontal lines. Diagonal lines are less subject to this strobing when the camera is panned and tilted. By extension a CCD array layed out at 45 degrees would perform in a similar manner. However if you took the Q67 to a bridge that had 45 degree truss elements, then you might get into trouble! So if you have a blocking filter that removes all fine features then this is not a problem, but this is not perhaps practical as who wants blurry images (as Graeme has pointed out).


It happens to be a visual artifact that humans in the physicaly constructed rectilinear world are going to come up against this time and time again. However if you shoot in the African Savanna with 3% lens distortion, then you will never experience this artifact.

But I don’t really think the 45 degree twist (of the SONY sensor) was designed to eliminate rectilinear dynamic aliasing artifacts. I think the 45 degree twist is just a necessity for speed and electronic chip architecture , design and physical layout.


I like David’s idea just to physically try these things out.

 

[I Bloom]

Eric, don’t get excited. They have to differentiate themselves somehow and when technology is pushed to the limits like this sometimes it is difficult to differentiate.

Guess what, I wouldn’t mind to have two boxes of Shreddies on the shelf in the kitchen.
Imagine, you wake up and you look on the shelf with two boxes of Shreddies side by side one square and one diamond.

Which one would you pick if you eat squares Shreddies for past 5 days?

I guess the point is that Q67 is a bayer pattern. Absolutely no difference. I wag the shame finger at anyone who thinks for a second that there might be some advantage or one orientation over another. Photosites aren't diamonds or squares, from the point of view of the recorded image they might as well be little circles.

A proper demosaic is always going to be better than a fast demosaic, again there is no question of that.

So the whole idea of even trying to market one against the other, by turning it 45 degrees, is either disingenuous, or shows they don't know what their doing.

Don't get me wrong. The F35 sounds like it's going to be quite a nice camera. It seems like they have a lot of selling points that don't involve mathemagical shenanigans.

IBloom

 

[Stuart English]

Mezzanine Codecs

Mezzanine Codecs

One thing that was mentioned was the need for mezzanine codecs for enabling reasonably cost effective 4K post.

Nice goal, but the "mezzanine codec" described in this presentation is 1920 x 1080 23.98 fps 10-bit 4:2:2 progressive,

or 1920 x 1080 59.94 fps 10-bit 4:2:2 interlace, at a data rate of 220Mb/s (~28MB/s ...similar to 4K RAW @ REDCODE 28 )

That is half normal HDCAM SR data rate, and similar to what already exists for both the ProRES and DNxHD editing codecs.

 

[M Most]

I guess the point is that Q67 is a bayer pattern. Absolutely no difference. I wag the shame finger at anyone who thinks for a second that there might be some advantage or one orientation over another. Photosites aren't diamonds or squares, from the point of view of the recorded image they might as well be little circles.

Except that they're not. In a Bayer sensor (and I would argue that the Q67 is NOT a Bayer sensor, but I'll get to that...) they are square photoreceptor sites arranged in a checkerboard grid. They are all the same size, and they are not circular. There are a number of reasons why both of these facts are significant. If you look at film, none of the receptors are the same size or shape (at least not predictably) and they are not arranged in a geometrical pattern, they are randomly dispersed. In fact, their varying size and shape is, in part, intentional in that it allows for good performance at various exposure levels within a typical image. There have, in fact, been some proposed electronic sensor designs based on the concept of photosites with different sizes and sensitivities within one chip (see Panavision's Dynamax proposal). Aliasing is, in part, a direct result of imposing a strict grid pattern on the photoreceptors.

As for the Bayer designation, I would argue that the notion that if it looks like a dog and smells like a dog it must be a dog does not really apply. One could describe a diamond as a square turned on its side, but it really isn't and most people don't refer to it that way. A Bayer pattern is, by definition, a vertical and horizontal grid with color filters laid out over that grid in a very specific way. In other words, it consists of squares. It does not consist of diamonds. That might seem a subtle distinction, but it may or may not be in the scheme of things. For one thing, depending on the approaches to reconstruction, a diamond pattern does apparently give you more of a contribution of the red and blue information on a per-green-pixel basis when compared to a pattern of squares. And while I don't disagree with Graeme's assertion that given a proper algorithm it's kind of a moot point, Graeme doesn't run Sony R&D and likely doesn't really know why they committed to their variation on the more traditional approach (or maybe he does, if so, I'd sure like to know that answer...). But apparently unlike some other posters here (NOT Graeme), I don't happen to believe that Sony is stupid, technically inept, or purely driven by evil marketing forces any more than other companies are. So I give them the benefit of a doubt until they deliver product that can be properly evaluated, the claims either proven or disproven, and the tradeoffs properly judged. But claiming that the Q67 approach is a Bayer pattern is not that dissimilar to saying that Redcode is JPEG2000. It's not. It's based on J2K, just as Q67 is based on Bayer, but that doesn't make it the same thing.

 

[Bruce Allen]

I may have missed this...
I'm certainly no sensor engineer, but I am curious. Why would rotating pixels 45 degrees make processing easier or quicker? Compared to handling standard rows and columns, it actually seems like it would be a bit more complex...unless of course it just processed along a 45 degree as well...in which case it's still doing the same old thing? :001_unsure:

You're not missing a thing, Michael. But some people in this thread seem to think that the Q67 approach will offer easier processing. I think they're assuming that the camera will do all demosaic processing onboard and pass off a 4K RGB file of some sort. Thus eliminating the need for complex debayer processing in post or edit. Either way, I don't see the advantage there... Or at least no advantage that wouldn't include a huge compromise.

As far as I understand it...

It's a combination of rotating things by 45 degrees, ALONG WITH the mandatory spec of pixel density so that you have exactly one green sensor per pixel. Which you don't have with Bayer.

Again, the whole point is so that you get "good enough" quality realtime 4K RGB output, easily, out of the camera.

Here's a naive algorithm for the top left pixel that lets you see how simple and easy it is:

With 18MP Q67:
G = green sample (there's one for each pixel)
R = (R sample at top left + R sample at bottom right) / 2
B = (B sample at bottom left + B sample at top right) / 2
(probably you'd want to sample some of the other R and G pixels, but you get the idea

Implemented in a device, you could just send the R and B pixel data through a simple, cheap, super-fast, fixed function, bilinear interpolation chip, to end up with the 4K output.

Memory requirement is just the original image (17.7M x 12bits = 26.6 MB), plus a 4K RGB result. Let's say it's 3840 x 2160 x 12bits x 3 channels... approx 35.6MB.

So total memory use = 62.2 MB

Or possibly they're storing internally as YUV 4:2:2? That'd cut down memory use even more. Since G is the largest component of the Y channel, and it is being sampled at 2x the rate of the other R and B channels, the sensor data would map to YUV extremely well, with very little processing involved.

Now, with 18MP Bayer naive algorithm:
G = G sample every 2nd pixel - otherwise interpolate
R = R sample every 4th pixel - otherwise interpolate
B = B sample every 4th pixel- otherwise interpolate

But in actual fact most Bayer algorithms are far more complex, as the above one doesn't get the max out of each pixel. PLUS you're doing the whole algorithm at 6K resolution, so it takes longer.

The result is a 6K RGB image (let's say 5760 x 3240 x 12 x 3) that needs to be stored in memory = 80MB.

Then you need to scale it down to 4K RGB and store THAT in memory = another 35.6MB.

So total memory use per frame= 142.2MB vs 62.2MB for Q67. EG more than double.

Computation cost is higher:
1x simple process (applied at 4K = eg to 8 million pixels)
VS:
1x more complex process (applied at 6K = eg to 18 million pixels)
plus a HQ scale from 6K to 4K

The other thing is: with Bayer, you are designing your OLPF to be a compromise between the G info and the R and B info. So you either:

- have the OLPF too high: lose G info because you're blurring to avoid any R or B aliasing
- have the OLPF in the middle: lose a bit of G info and get a bit of R and B aliasing
- have the OLPF too low: get all of the G info but get heavy R and B aliasing

If you know you're doing a Q67 decode, you can choose a more balanced OLPF. YOU KNOW that the resolution won't be as good as Bayer for the G channel, but that's okay, because if you had chosen the OLPF correctly to avoid aliasing in R and B channels, you would have lost a lot of that data anyway.

OF COURSE it would be better to have 6K Bayer sensor and then interpolate down later... EG what everyone is comparing this to. But Sony was never considering that - because REALTIME 4K OUTPUT was a design requirement.

Fact is, people expect different things from Sony than they do from RED because of Sony's broadcast, realtime tradition.

You can buy a $50K Red camera package... and somehow nobody expects it to be able to actually output 4K from the camera.

If in 5 years' time you told Joe Consumer:
"Hey man - in order to plug your new $3000 Sony 4K camcorder into your new Sony 4K TV you need a $5000 SonyRocket plugged in... he'd throw it across the room at you."
Same thing applies for broadcast. Broadcast guys will expect a 4K camera with 4K outputs.

Bruce Allen
www.boacinema.com

 

[Graeme Nattress]

"a diamond pattern does apparently give you more of a contribution of the red and blue information on a per-green-pixel basis when compared to a pattern of squares" - how and why? Sony don't say that. I don't say that.

I don't feel very gracious about giving people the benefit of the doubt. I had enough people, even those I knew and trusted basically telling me that no way on earth the RED one could work as described.

Most of the time, I really like to measure things, but with enough experience and enough knowledge of math you can make fairly accurate predictions, and for those I go with measured data on the traditional Bayer and gave Q67 a very positively biassed set of numbers and it still doesn't make sense.

Graeme

 

[Bruce Allen]

I don't feel very gracious about giving people the benefit of the doubt. I had enough people, even those I knew and trusted basically telling me that no way on earth the RED one could work as described.

Most of the time, I really like to measure things, but with enough experience and enough knowledge of math you can make fairly accurate predictions, and for those I go with measured data on the traditional Bayer and gave Q67 a very positively biassed set of numbers and it still doesn't make sense.

Graeme, are you actually, really flummoxed about why Sony went with this?

Again I ask - if you:
- had to make a camera with decent realtime 4K output
- if you didn't have much processing power to play around with
- and if you had a fantastic sensor fabrication facility

What is wrong with Q67 as a solution? Where Q67 is NOT JUST turning bayer on its side... but speccing the chip so that there's exactly 4K green pixels for a 4K output. It allows for a good OLPF compromise.

Doing a 5792 x 3258 Bayer is NOT an option for Sony for computational reasons. So what do you propose instead? That a 4K Bayer is better than their current suggestion? You know it's not.

At the moment you're comparing apples to oranges - like saying "My chess algorithm beats theirs"... Without mentioning that their algorithm is designed for speed chess.

Bruce Allen
www.boacinema.com

 

[Zach Gray]

OF COURSE it would be better to have 6K Bayer sensor and then interpolate down later... EG what everyone is comparing this to. But Sony was never considering that - because REALTIME 4K OUTPUT was a design requirement.

Fact is, people expect different things from Sony than they do from RED because of Sony's broadcast, realtime tradition.

...

Same thing applies for broadcast. Broadcast guys will expect a 4K camera with 4K outputs.

Bruce Allen
www.boacinema.com

This makes the most sense of any justification I've read! It's been marketed as a better design, when in reality it's just a faster design... realtime broadcast spec.

 

[Graeme Nattress]

Bruce, I know what I'd do alright! I'd just not say publicly....

Graeme

 

[Bill Anderson]

Well, a square turned on its "side" is still a square. ( ;
Seriously though, a "diamond", as I understand it, is a lozenge, and not a square set at a different orientation. A diamond would be a different kettle of fish. Sony has said or shown nothing that would make one believe that their Q67 is not a Bayer pattern skewed - and if we are to avoid "if it smells like a dog", or even "a zebra standing on its hind legs is still a zebra" then: if I turn my Bayer sensor camera at an angle, is it no longer a bayer pattern sensor - even if it performs slightly differently????
Has Sony shown that it is redistributing the typically bayer color assigned pixels... ???

 

[Eric Lange]

Pompi's Hallway

Pompi's Hallway

In the spirit of the Open University…think about this…

Imagine you have a large hall way, and you decide that you want to go all Roman and lay yourself a beautiful mosaic floor.

The hall way is 2 meter by 5 meter, in other words 200 by 500 cms. Being Roman you have many thousands of tesserae on hand (the proper name and article for the small square blocks that go to make up a mosaic). You decide to go all post modern and have red green and blue tessera, but you want two green ones for every blue and red one. You decide to lay the floor out in a rectilinear grid parallel to the walls, with the green tiles making diagonal stripes and alternate the corners with either a blue one or a red one. Each block is 1cm by 1cm by 1cm or 10mm in each dimension. Therefore you deduce that you need a total of 100,000 tesserae in the following proportions 2:1:1. Therefore 50,000 green ones, 25,000 blue ones and 25,000 red ones.

After some serious labor and the grout has set, your brother in law Pompi comes round, and is so impressed by the floor he says hey I have a 2 meter by 5 meter hallway also, can you make me a mosaic floor too. Except Pompi is a real control freak and instead just to be different he says he wants all the tiles to lie at 45 degrees to the wall but in the same ratios of two green to one blue to one red. He says instead he wants the centers of the green tiles/tesserae to be parallel to the walls. Ok not too difficult..

QUESTION 1.

How many RED, GREEN and BLUE tesserae does Pompi need for his floor with diagonally arranged tiles, given that it is the exact same area? (And yes you count the ones you had to cut in half corner to corner for each tile that meets the wall and each tessera still measures 10mm in all dimensions).

 

[M Most]

Well, a square turned on its "side" is still a square. ( ;
Seriously though, a "diamond", as I understand it, is a lozenge, and not a square set at a different orientation. A diamond would be a different kettle of fish. Sony has said or shown nothing that would make one believe that their Q67 is not a Bayer pattern skewed - and if we are to avoid "if it smells like a dog", or even "a zebra standing on its hind legs is still a zebra" then: if I turn my Bayer sensor camera at an angle, is it no longer a bayer pattern sensor - even if it performs slightly differently????
Has Sony shown that it is redistributing the typically bayer color assigned pixels... ???

The design has nothing to do with the sensor orientation vis a vis the scene it's photographing. It has to do with RGB image reconstruction, and how the information is read out and processed.

I have to say it's a bit tiring to see the constant disdain and disregard for any alternate approaches here, especially without any real visual evidence to support either view. And it's also a bit frustrating to see the continuing lack of consideration for markets and purposes other than digital cinema, in spite of the fact that many of Red's own people do continue to point out the importance of such markets for things like Red Ray and the Scarlet line. As Bruce pointed out, products are designed for particular performance characteristics to meet particular market needs. Red One isn't a particularly practical or appropriate camera for live HD broadcast. And the Sony F900 derived products are not particularly optimal for digital cinema. So what?

 

[M Most]

they may or may not be. but we know for sure some are, and i'm not talking about companies. let's give it a rest.

shoot your camera at a 45 degree angle and be done with it, i don't want to hear anything else until i see real tests and a deliverable.

Paul, I don't know what you have against me or why you have to get personal any time I post something here, but I really resent your implication. I try pretty hard not to insult the intelligence of anyone here, including you. I would ask that you do the same.

I reserve the right to refuse replies to anyone with a bad attitude. :emote_hippie:

 

[Eric Lange]

Except that they're not.

....

As for the Bayer designation, I would argue that the notion that if it looks like a dog and smells like a dog it must be a dog does not really apply. thing.

A dog rotated 45 degrees still looks and smells the same to me (just don’t squeeze too hard).

“Does Your Dog Bite…”

www.youtube.com/watch?v=SXn2QVipK2o

 

[Bill Anderson]

Don't forget, Sony hasn't qualified their Q67 as being useful for only A to B- they have unabashedly put it up against red and alexa to point out their flaws/shotcomings/failings/inferiority...

As far as Sony bashing is concerned, things can get a bit hot for sure, but that (and shape semantics) aside, I'd still like to learn why the Q67 is a superior approach.
And this is not some newcomer we're dealing with here, this is the Sony corporation: Jim and Graeme have taken a few body blows in their time, and still more to come, no doubt, let's see how Sony does. Gloves off. ( :

 

[Bruce Allen]

Don't forget, Sony hasn't qualified their Q67 as being useful for only A to B- they have unabashedly put it up against red and alexa to point out their flaws/shotcomings/failings/inferiority...

As far as Sony bashing is concerned, things can get a bit hot for sure, but that (and shape semantics) aside, I'd still like to learn why the Q67 is a superior approach.

It's not a superior approach quality-wise - IF you're comparing to a Bayer sensor of the same resolution and IF the Bayer sensor is attached to the powerful, expensive demosaic and scaler chips it needs. Two big IFs.

Sony's new sensor likely beats the current Alexa and Red One in terms of resolution and capabilities (realtime 4K output, for example). It'll likely have worse quality than Epic but better broadcast capabilities (realtime 4K output) - until Red makes a Rocket add-on module.

Note that in this thread Graeme hasn't been comparing the Sony design with a Red One, but VS a hypothetical 6K Bayer camera.

A dog rotated 45 degrees still looks and smells the same to me (just don’t squeeze too hard).
“Does Your Dog Bite…”
www.youtube.com/watch?v=SXn2QVipK2o

Love Peter Sellers. So, did you bother to actually read what Mike or I posted? It's about designing a sensor around easy-to-decode, "good-enough quality" 4K realtime output.

For example, in broadcast situations, realtime 4K output might be a nice thing to do on-camera. So that you can watch sports AS IT HAPPENS, for example. Radical concept, I know.

Anyone actually want to read my posts and come back with a well reasoned answer about why Q67 is bad from a computational or OLPF design perspective - with the goal being 4K realtime output for broadcast?

The fact is that Red doesn't complete the 4K puzzle at the moment. The moment we can get an easy 4K broadcast pipeline is the moment 4K TVs stop costing stratospheric prices.

Sony's (Q67-based) sensor seems to be a good compromise to solve those problems. Epic with a built-in RED ROCKET-style module sounds like another. But instead of talking about that, everyone is running around comparing Q67 to a 6K Bayer sensor, making stupid diamond analogies and completely missing the point.

Bruce Allen
www.boacinema.com

 

[Graeme Nattress]

Bruce, the point is that Q67 has been aggressively marketed as a sensor design against us "Bayer traditionalists" with some real slight of hand (doubling the pixel count) to "prove" it's a superior design, whereas simple analysis shows otherwise. That's why we compare same number of pixels because their argument is about a sensor design, not a camera. When they have a camera, we can compare that with a current Epic and see how things line up.

As for Q67 simplifying processing, it does so only at the expense of doubling the number of pixels needing processing - I don't see it as a free lunch in that regard.

Graeme

 

[Michael Hastings]

The design has nothing to do with the sensor orientation vis a vis the scene it's photographing. It has to do with RGB image reconstruction, and how the information is read out and processed.

I have to say it's a bit tiring to see the constant disdain and disregard for any alternate approaches here, especially without any real visual evidence to support either view. And it's also a bit frustrating to see the continuing lack of consideration for markets and purposes other than digital cinema, in spite of the fact that many of Red's own people do continue to point out the importance of such markets for things like Red Ray and the Scarlet line. As Bruce pointed out, products are designed for particular performance characteristics to meet particular market needs. Red One isn't a particularly practical or appropriate camera for live HD broadcast. And the Sony F900 derived products are not particularly optimal for digital cinema. So what?

Mike I for one really appreciate your input.

And I too, get tired of the disdain/disregard factor - and I'm a person that is all in with Red - I no longer own any cameras from Sony, panasonic, canon or Nikon worth more than $1200 retail. But I'm a little concerned that the tone here on Reduser affects the perception of the camera and those who use it - regardless of technical merit - and perception is reality in the rental market and even in the freelance market where, right or wrong, many hires are based on your camera package.

In another thread someone mentioned that many DPs are against Red and rob ruffo seemed to be befuddled and asked why. Well it's because being professional means being aware of and knowledgeable of the merits of alternative approaches, AND to be gracious within a competitive environment. This is somewhat lacking with some redusers. It may just be a vocal few - but it's an issue.

I can't sort this one out based only on presentations, and it makes no difference to my foreseeable purchases, but Sony has been on the forefront of virtually every electronic motion picture advancement during my 30 year career and I would tend to give them the benefit of a look before I dismiss them out of hand.