New Sensor Tech in EOS 7D Mark II [CR2]

[LetTheRightLensIn]

Right.
What would be really cool to see is some sort of hardware level binning process that maintains the integrity of the RAW file.

Half the reason I'm so anxious for super high resolution cameras is that I haven't been terribly impressed with the image quality off my 5D2. That nasty AA filter (which I'm pretty sure is especially bad on the 5D2) effectively cuts resolution in half. When I first saw my pictures on a decent 4MP monitor I was amazed at how little detail loss there was vs. looking at the image zoomed to 100%. My bet is that a good 4K (8MP) monitor is going to display your images with just as much detail as a high quality print... Because the detail actually isn't there in the first place.

One option is just quadrupling resolution and getting rid of the AA filter (which I'm actually fine with), but if they could bin the full per-pixel RGB signal on the sensor it should effectively deal with moire, and we get to keep our current file size, and it should produce an actual 20MP image instead of the blurred out fake we currently end up with.

The last thing I really want to see is the integration of clear microlenses. Even the heavily faded green pixels that we have right now still block a lot of light. Given how advanced interpolation is I doubt that eliminating the colour value for one of the pixels would have a significant impact on image quality.

Sorry, but that (bolded) is such a ludicrous, laughable comment, I'm just flabbergasted. An AA filter DOES NOT cut resolution "in half". That is blowing things SO FAR out of proportion it may be one of the most ludicrous things I've read on these forums. OLPFs, optical low pass filters, are designed to affect high frequencies only, and only around the nyquist limit at that. You lose a TINY amount of resolution...but it doesn't matter, because the "resolution" your losing just contains nonsense anyway. OLPFs blur very high frequency data that nearly or exactly matches the spatial frequency of the sensor's pixels just enough such that they the information doesn't alias. That's it. Aliased information is a REAL loss of information. Technically speaking, OLPFs PRESERVE information...they save information that can be saved, and discard information that cannot be correctly interpreted by the sensor anyway. On top of that, a very light application of unsharp masking can effectively reverse the blurring, and improve the resolution of that high frequency data, without actually bringing back all the nonsense.

Quadrupling resolution and removing the AA filter is only an option if your lenses cannot resolve that much detail. With the resolving power of Canon's current lens lineup at faster apertures, I'm not so sure that cutting pixels into quarters is actually enough to avoid any kind of aliasing. At narrower apertures, like f/8, diffraction already blurs information enough that it can't alias, but that's a really narrow aperture for a lot of work, not everyone uses it. There are very few applications where removal of an AA filter will not cause aliasing of some kind, and pretty much anything artificial is going to have repeating patterns that, depending on distance to camera, can create interference patterns (moire).

This whole "Remove the AA filter" craze is just that...a craze. It's a "thing" Nikon started doing to be different, to get some "wows", and maybe bring in some more customers. Ironically, given that removal of an AA filter is really NOT a good thing...it's worked. Nikon's marketing tactics have sucked in a whole lot of gullibles who don't really know what an AA filter does or how it works, or how to work WITH it, and now we have a whole army of "photographers" who want AA filters removed from all cameras. Personally, I REALLY, TRULY, HONESTLY DO NOT want Canon to remove the AA filter. It is NECESSARY, it PRESERVES preservable data and eliminates useless data, and I LIKE THAT.

And anything that is lost? It's MINIMAL. In the grand scheme of how much resolution you have...you maybe lose a percent or two of really high frequency information...but you really don't have that information anyway because it is similar in frequency to noise...so again, moot.

Given that the filter makes it physically impossible to have a repeating pattern of stripes the same frequency as the pixel grid, so that you cannot have a perfect transition of black pixels to white, I'd say that is cutting resolution in half. That is, compared to some magical thing that accurately reads the full RGB spectrum on each pixel.

You are right about the necessity of the AA filter though.
I was thinking that if the interpolation algorithm only sampled each pixel within a specific cluster of four pixels and not every pixel around it that it would solve the moire problem. Really that would just give you different colour banding instead.
Now, if we added a second layer of microlenses on top of the first to direct light only at individual groups of pixels, that would guarantee the full RGB read on each cluster, and allow hard transitions...

On second thought I guess that sounds a little excessive just to gain the ability to have large pixels with a hard transition instead of twice as many pixels with a row of grey pixels that's half as big. You can bin the smaller pixels with a normal AA filter just the same, we just need a way of doing that without destroying the flexibility of RAW (otherwise I assume people would have been using compressed formats a long time ago).

it's not cutting it in half

 

[9VIII]

Right.
What would be really cool to see is some sort of hardware level binning process that maintains the integrity of the RAW file.

Half the reason I'm so anxious for super high resolution cameras is that I haven't been terribly impressed with the image quality off my 5D2. That nasty AA filter (which I'm pretty sure is especially bad on the 5D2) effectively cuts resolution in half. When I first saw my pictures on a decent 4MP monitor I was amazed at how little detail loss there was vs. looking at the image zoomed to 100%. My bet is that a good 4K (8MP) monitor is going to display your images with just as much detail as a high quality print... Because the detail actually isn't there in the first place.

One option is just quadrupling resolution and getting rid of the AA filter (which I'm actually fine with), but if they could bin the full per-pixel RGB signal on the sensor it should effectively deal with moire, and we get to keep our current file size, and it should produce an actual 20MP image instead of the blurred out fake we currently end up with.

The last thing I really want to see is the integration of clear microlenses. Even the heavily faded green pixels that we have right now still block a lot of light. Given how advanced interpolation is I doubt that eliminating the colour value for one of the pixels would have a significant impact on image quality.

Sorry, but that (bolded) is such a ludicrous, laughable comment, I'm just flabbergasted. An AA filter DOES NOT cut resolution "in half". That is blowing things SO FAR out of proportion it may be one of the most ludicrous things I've read on these forums. OLPFs, optical low pass filters, are designed to affect high frequencies only, and only around the nyquist limit at that. You lose a TINY amount of resolution...but it doesn't matter, because the "resolution" your losing just contains nonsense anyway. OLPFs blur very high frequency data that nearly or exactly matches the spatial frequency of the sensor's pixels just enough such that they the information doesn't alias. That's it. Aliased information is a REAL loss of information. Technically speaking, OLPFs PRESERVE information...they save information that can be saved, and discard information that cannot be correctly interpreted by the sensor anyway. On top of that, a very light application of unsharp masking can effectively reverse the blurring, and improve the resolution of that high frequency data, without actually bringing back all the nonsense.

Quadrupling resolution and removing the AA filter is only an option if your lenses cannot resolve that much detail. With the resolving power of Canon's current lens lineup at faster apertures, I'm not so sure that cutting pixels into quarters is actually enough to avoid any kind of aliasing. At narrower apertures, like f/8, diffraction already blurs information enough that it can't alias, but that's a really narrow aperture for a lot of work, not everyone uses it. There are very few applications where removal of an AA filter will not cause aliasing of some kind, and pretty much anything artificial is going to have repeating patterns that, depending on distance to camera, can create interference patterns (moire).

This whole "Remove the AA filter" craze is just that...a craze. It's a "thing" Nikon started doing to be different, to get some "wows", and maybe bring in some more customers. Ironically, given that removal of an AA filter is really NOT a good thing...it's worked. Nikon's marketing tactics have sucked in a whole lot of gullibles who don't really know what an AA filter does or how it works, or how to work WITH it, and now we have a whole army of "photographers" who want AA filters removed from all cameras. Personally, I REALLY, TRULY, HONESTLY DO NOT want Canon to remove the AA filter. It is NECESSARY, it PRESERVES preservable data and eliminates useless data, and I LIKE THAT.

And anything that is lost? It's MINIMAL. In the grand scheme of how much resolution you have...you maybe lose a percent or two of really high frequency information...but you really don't have that information anyway because it is similar in frequency to noise...so again, moot.

Given that the filter makes it physically impossible to have a repeating pattern of stripes the same frequency as the pixel grid, so that you cannot have a perfect transition of black pixels to white, I'd say that is cutting resolution in half. That is, compared to some magical thing that accurately reads the full RGB spectrum on each pixel.

You are right about the necessity of the AA filter though.
I was thinking that if the interpolation algorithm only sampled each pixel within a specific cluster of four pixels and not every pixel around it that it would solve the moire problem. Really that would just give you different colour banding instead.
Now, if we added a second layer of microlenses on top of the first to direct light only at individual groups of pixels, that would guarantee the full RGB read on each cluster, and allow hard transitions...

On second thought I guess that sounds a little excessive just to gain the ability to have large pixels with a hard transition instead of twice as many pixels with a row of grey pixels that's half as big. You can bin the smaller pixels with a normal AA filter just the same, we just need a way of doing that without destroying the flexibility of RAW (otherwise I assume people would have been using compressed formats a long time ago).

it's not cutting it in half

Yes and no?

In the theoretical perfect transition, with no AA filter you get a "black-white" transition, and then you have a "black-grey-grey-white" transition with the AA filter.
Ideally with the AA filter the line would land directly in the middle of the pixels, which would actually produce exactly the same result as having no AA filter, a "white-grey-black" transition. But that is equally improbable as the perfect pixel transition, you're practically going to end up with a transition going from white, to two pixels of varying shades of grey, to black. Whereas without your worst case scenario is one line of grey pixels.

So the best case scenario with an AA filter is somewhere between a 50%-100% increase in blur. Where without you go somewhere from a 0% to 50% increase in blur. On a theoretical perfect line.
Yes, it's nitpicking, but that's what makes the forums so much fun.

 

[bdunbar79]

What do you all think are the chances, that the 7D II will surpass the 5D mark iii/1DX on certain things? Such as IQ, ISO, DR or other things?

IMO, the 7DII will at best match the ISO/noise performance of the 5DIII.
And if Canon has finally decided to implement on-chip analog-to-digital conversion (ADC),
DR at low ISO could be better than on the FF cameras.
That's about it, though, in terms of IQ.

Also, just like with the 7D, we might see some features on the 7DII that will later make it
into the higher end cameras.

Overall, it's hard to imagine that the 7DII will offer much more from what you can get
already with the 5DIII today (except for higher frame rate and more pixels per duck).

Not if it's not FF. No way.

 

[LetTheRightLensIn]

FWIW, if the photodiode filter arrangement is not the classic Bayer type, but uses some other arrangement (Fuji being the main example), the major software companies, in addition to the DPP in-house software team, will need a bit of time to implement the new filter arrangement into their RAW converters. On the off chance that someone is hoping for a Foveon-ish sensor, the developers will need more than "a bit" of time because the algorithms are a lot different. There's only one non-Sigma RAW developer out there that can use x3f files, Iridient Developer.

the algorithms are much different, but also much, much simpler

 

[9VIII]

As for the double layer of microlenses...sure, you could read a full RGBG 2x2 pixel "quad" and have "full color resolution". Problem is, that LITERALLY halves your luminance spatial resolution...

Thus you start with an 80MP sensor to get a nice 20MP image.

BTW, what your describing is called super-pixel debayering. That, too, is a common option in astrophotography image stacking...instead of basic or AHD debayering, you usually have the option to either super-pixel debayer, or "drizzle" (which, if you have enough subs...such as a couple hundred...is a means of achieving superresolution, and can increase your output image resolution by two to three fold.) You don't even need another microlens layer to do super-pixel debayering...you could use a tool like Iris or maybe even DarkTable/RawThearapy, to do it on any image you want.

Finally, even if you do super-pixel debayering, your not going to ever have "hard edges". Statistically speaking, the chances if a white/black line pattern you wish to photograph perfectly lining up with your pixels, regardless of how large or small they are, is so excessively remote that it is statistically impossible. Not in any real-world situation. You might be able to build some kind of contraption and AI software to eventually achieve it, but that is well beyond the realm of practicality. If you remove the AA filter, use super-pixel debayering, you might have larger pixels with full color fidelity...but your going to have a massive amount of aliasing. Those white and black lines would have some nasty stair-stepped edges, they would just look atrocious.

Wow, it looks like superpixel debayering (http://pixinsight.com/doc/tools/Debayer/Debayer.html) is exactly what I'm after. Make a 128MP sensor and use superpixel debayering and you'll have a nice compact, super accurate 32MP image.
Again, really, I'm fine with just shooting on a 128MP sensor and dealing with 100MB+ RAW files, the trick is to get a similar result in a format that's going to be acceptable to the majority of photographers who refuse to deal with large file sizes.

As long as your final image is around 32MP I don't think people are going to notice the stair stepping, unless you're standing right next to something like a 40" high quality print.

 

[Don Haines]

What do you all think are the chances, that the 7D II will surpass the 5D mark iii/1DX on certain things? Such as IQ, ISO, DR or other things?

The 7DII will surpass the 5DIII/1D X in viewfinder magnification, and the 5DIII in frame rate...that's pretty much it.

There also might be several small items with the 7DII like wifi, GPS or a built-in-flash. You can argue about whether you need them, but the 7DII will surpass the other two models on that.

It will be cheaper, too. That's probably the biggest benefit!

I expect it will also have a touch-screen.

Wowzers....killer new feature..... :-\

Hey, don't knock it. After all, changing settings with your nose is better than butt-dialing.

GREAT! Yet another image I will never get out of my head..... Neuro dialing a phone with his butt.......

 

[ScottyP]

I'll toss it out there.......


APS-H?

 

[Dylan777]

What do you all think are the chances, that the 7D II will surpass the 5D mark iii/1DX on certain things? Such as IQ, ISO, DR or other things?

IMO, the 7DII will at best match the ISO/noise performance of the 5DIII.
And if Canon has finally decided to implement on-chip analog-to-digital conversion (ADC),
DR at low ISO could be better than on the FF cameras.
That's about it, though, in terms of IQ.

Also, just like with the 7D, we might see some features on the 7DII that will later make it
into the higher end cameras.

Overall, it's hard to imagine that the 7DII will offer much more from what you can get
already with the 5DIII today (except for higher frame rate and more pixels per duck).

Wishful thinking and I hope you right. The longest lens I have is 400mm f2.8 IS II, I don't mind adding 7D II(x1.6) if that the case

 

[jrista]

As for the double layer of microlenses...sure, you could read a full RGBG 2x2 pixel "quad" and have "full color resolution". Problem is, that LITERALLY halves your luminance spatial resolution...

Thus you start with an 80MP sensor to get a nice 20MP image.

No, that is fundamentally incorrect. You start with a 20mp sensor, which has 40mp PHOTODIODES. The two are not the same. Pixels have photodiodes, but photodiodes are not pixels. Pixels are far more complex than photodiodes. DPAF simply splits the single photodiode for each pixel, and adds activate wiring for both. That's it. It is not the same as increasing the megapixel count of the sensor.

And, once again...I have to point out. There is no such thing as QPAF. The notion that Canon has QPAF is the result of someone seeing something they did not understand. Canon does not have QPAF. Their additional post-DPAF patents do not indicate they have QPAF technology yet...however there have been improvements to DPAF.

BTW, what your describing is called super-pixel debayering. That, too, is a common option in astrophotography image stacking...instead of basic or AHD debayering, you usually have the option to either super-pixel debayer, or "drizzle" (which, if you have enough subs...such as a couple hundred...is a means of achieving superresolution, and can increase your output image resolution by two to three fold.) You don't even need another microlens layer to do super-pixel debayering...you could use a tool like Iris or maybe even DarkTable/RawThearapy, to do it on any image you want.

Finally, even if you do super-pixel debayering, your not going to ever have "hard edges". Statistically speaking, the chances if a white/black line pattern you wish to photograph perfectly lining up with your pixels, regardless of how large or small they are, is so excessively remote that it is statistically impossible. Not in any real-world situation. You might be able to build some kind of contraption and AI software to eventually achieve it, but that is well beyond the realm of practicality. If you remove the AA filter, use super-pixel debayering, you might have larger pixels with full color fidelity...but your going to have a massive amount of aliasing. Those white and black lines would have some nasty stair-stepped edges, they would just look atrocious.

Wow, it looks like superpixel debayering (http://pixinsight.com/doc/tools/Debayer/Debayer.html) is exactly what I'm after. Make a 128MP sensor and use superpixel debayering and you'll have a nice compact, super accurate 32MP image.
Again, really, I'm fine with just shooting on a 128MP sensor and dealing with 100MB+ RAW files, the trick is to get a similar result in a format that's going to be acceptable to the majority of photographers who refuse to deal with large file sizes.

As long as your final image is around 32MP I don't think people are going to notice the stair stepping, unless you're standing right next to something like a 40" high quality print.

Well, someday we may have 128mp sensors...but that is REALLY a LONG way off. DPAF technology, or any derivation thereof, isn't going to make that happen any sooner.

 

[jrista]

Yes and no?

In the theoretical perfect transition, with no AA filter you get a "black-white" transition, and then you have a "black-grey-grey-white" transition with the AA filter.
Ideally with the AA filter the line would land directly in the middle of the pixels, which would actually produce exactly the same result as having no AA filter, a "white-grey-black" transition. But that is equally improbable as the perfect pixel transition, you're practically going to end up with a transition going from white, to two pixels of varying shades of grey, to black. Whereas without your worst case scenario is one line of grey pixels.

So the best case scenario with an AA filter is somewhere between a 50%-100% increase in blur. Where without you go somewhere from a 0% to 50% increase in blur. On a theoretical perfect line.
Yes, it's nitpicking, but that's what makes the forums so much fun.

Your talking about the scientifically ideal situation. Those only exist in text books. They don't exist in reality, not with the countless other factors that go into resolving an image accounted for. That would be like saying that you could create the ideal frictionless surface often referred to in physics text books. You can greatly reduce the coefficient of friction, but you cannot eliminate it. You cannot actually achieve the perfectly ideal.

So there is the theory of resolving line pairs, and then there is the 100% perfectly ideal exemplar. Yes, in the ideal exemplar case, theoretically you could line up black and white lines perfectly on top of rows of pixels, and they would end up perfectly sharp. Perfect is unattainable.

When you account for other factors, such as the statistical improbability that you would EVER be able to line up alternating white and black lines perfectly on the sensor, the difference isn't blur...it's aliasing. You either end up with aliased results, which means you have "nonsense" information...or VERY SLIGHTLY blurred results for high frequency oscillations. It really isn't even blurring, it's frequency stretching or spreading, which effectively stretches high frequencies and makes them a slightly lower frequency, which actually represents the real information much more accurately than the nonsense. That is not a reduction in resolution, it's the elimination of useless data. Anti-aliasing is not designed to destroy information...it is actually designed to PRESERVE information, by throwing away what you cannot resolve accurately anyway.

The removal of an AA filter does not mean your producing more accurate images. Your producing less accurate images that have higher acutance. That's it. Thing about acutance is...it's easy to replicate, to the small degree necessary at high frequencies...with software. A simple unsharp mask will improve the acutance of an image taken with a camera that has an AA filter. The only difference between the two images at that point is that the anti-aliased image is accurate AND sharp, where as the aliased image is sharp but not accurate. There is really no benefit to removal of the AA filter unless your imaging highly random information. There are very few things like that. Landscapes come to mind as one of the primary, and very few, situations where removal of an AA filter could potentially be useful. I wouldn't even say macro photography would be better without an AA filter...when you magnify small subjects so significantly, there tends to be a LOT of high frequency data, and you would be surprised how often there are repeating patterns at the microscopic scale. Even without repeating patterns, certain natural features, such as the cells of an insect eye, end up looking more jagged and harsh than they do when you use a camera with an AA filter.

Sharpness isn't the supreme indicator of IQ. Too much sharpness is often the hallmark of significant overprocessing...a slight amount of softening of very high spatial frequencies is usually the hallmark of a skilled processor. Some of THE BEST landscape photography I admire the most has a very soft aesthetic, with a specific amount of slightly lower contrast in the high frequencies. These kinds of landscapes are the ones that really stand out from the throngs of landscape photos as being exceptional.

 

[x-vision]

No, that is fundamentally incorrect. You start with a 20mp sensor, which has 40mp PHOTODIODES.

Jrista, you are just assuming that Canon's dual-pixel tech is in fact a dual-photodiode tech.
My assumption is that it's already a quad-photodiode tech - and it's equally valid, as neither
one us has info on the actual implementation.

In general, before making any claims for photodiodes and pixels, consider the following:
A 'classic' pixel design has a photodiode plus three transistors (you can read about it on Wikipedia):

• a reset transistor for resetting the photodiode voltage
• a source-follower transistor for signal amplification
• a row select transistor

So, one definition of a pixel is a photodiode with three transistors.

The thing is, to improve fill factor and for other design considerations, modern sensors are using transistor sharing.
That is, a single set of the 'classic' transistors is shared between multiple phododiodes.

Transistor sharing is widely used in small sensors.
In the case of these sensors, though, each photodiode has its own microlens.
Thus, the photodiode is the pixel in these designs.

In short, depending on the implementation, a photodiode and a pixel could mean the same thing.

Canon's 'dual-pixel' tech is assumed to be based on a shared-transistor design.
That is, it is a multi-photodiode design.
But since in a shared-transistor design photodiodes are effectively equivalent to pixels (as explained),
Canon's tech could be called multi-pixel design as well.

So, you can stop correcting people who use dual/quad-pixel terminology, as these could in fact be used interchangeably.
The line between between a pixel and a photodiode is blurred in shared-pixel designs.
And the fact that the two photodiodes are read independently for auto-focus further
indicates that these could very well be independent pixels - if they didn't share the
same microlens and color filter.

Also, your claim that there are exactly TWO PHOTODIODES (and that's it!) is not based on fact.
We don't know for sure if Canon's design is a dual-pixel design (your assumption) or a quad-pixel design
(my assumption).

Canon's marking is selling it as a 'dual-pixel' tech likely because it's easier this way to communicate
the concept to the general public.
But we don't know for a fact what the actual implementation is.

So, your TWO PHOTODIODES claim is based on marketing materials, really.
If I were you, I wouldn't put too much weight into these 8).

My assumption for a quad-pixel design is based on simple geometry.
If there are just two photodiodes per pixel, these photodiodes need to be rectangular.
This would be uncommon - if not even a first in the industry.
But with a quad-pixel design, the photodiodes are square just like in any other sensor.

Considering the potential future advantages of a quad-pixel design (e.g. for a non-Bayer sensor),
I'd speculate that Canon would have invested in a quad-pixel design from the start - rather than
designing rectangular photodiodes that later would need to be made square anyway.

Just a speculation, of course - but based on some informed assumptions.

 

[Orangutan]

Jrista, you are just assuming that Canon's dual-pixel tech is in fact a dual-photodiode tech.

See: http://www.usa.canon.com/cusa/consumer/products/cameras/standard_display/daf_technology

Each pixel on the EOS 70D camera's sensor consists of two independent photodiodes that function both as imaging points and as individual phase-difference AF sensors.

It may be an assumption, but it's based on published material. What's the support for your assumption?

 

[x-vision]

... but it's based on published material.

Right. But that's still marketing materials.

 

[Orangutan]

... but it's based on published material.

Right. But that's still marketing materials.

They literally say two "photodiodes," not "photosites," or merely "pixels," but "photodiodes." That's very specific and technical (the typical consumer has little to no idea what a photodiode is). Why would they say two when it's really four? From the marketing perspective, four is better than two.

Magic 8-Ball says "All signs point to 2 photodiodes." You still might be right, but the bulk of evidence is against you.

 

[x-vision]

Magic 8-Ball says "All signs point to 2 photodiodes."

Fair enough.

You still might be right, but the bulk of evidence is against you.

Heh. All the evidence is coming from one source - Canon.
If it corroborated by at least one other source, I maybe wouldn't have argued.
But for now we either accept what Canon is saying ... or not 8).

 

[jrista]

No, that is fundamentally incorrect. You start with a 20mp sensor, which has 40mp PHOTODIODES.

Jrista, you are just assuming that Canon's dual-pixel tech is in fact a dual-photodiode tech.
My assumption is that it's already a quad-photodiode tech - and it's equally valid, as neither
one us has info on the actual implementation.

Sorry, but I am NOT assuming. I've actually read Canon's own patents. Those patents describe a system where the photodiodes for each pixel have been divided in half. This stuff isn't a mystery. Patentes are ESSENTIAL for the protection of intellectual property. Canon has been filing patents for DPAF for quite some time, a couple of years at least now, with the most recent ones being near the end of last year.

My assertions are based on concrete fact as described by the DPAF engineers at Canon themselves. Your assumptions are just that, assumptions based on an extremely TINY image posted on the ChipWorks page of the BACKSIDE of some sensor, an image which you have gravely misinterpreted, and an image we all can only assume is even of a Canon sensor, let alone one with DPAF technology (although it certainly is not of a Canon sensor with QPAF technology...since such a sensor doesn't exist yet.)

In general, before making any claims for photodiodes and pixels, consider the following:
A 'classic' pixel design has a photodiode plus three transistors (you can read about it on Wikipedia):

• a reset transistor for resetting the photodiode voltage
• a source-follower transistor for signal amplification
• a row select transistor

So, one definition of a pixel is a photodiode with three transistors.

Sure. An extremely basic kind of "pixel" that you might find in an entry level course on image sensor design. Modern sensors often have a lot more logic than that per pixel. That logic usually involves some level of noise reduction, potentially charge bucketing for global shutter sensors, anti-blooming gates and shift registers in CCDs, extra logic to allow the selection of which photodiode to read in shared-pixel designs (which most smaller-pixel sensor designs are these days), etc.

The thing is, to improve fill factor and for other design considerations, modern sensors are using transistor sharing.
That is, a single set of the 'classic' transistors is shared between multiple phododiodes.

Transistor sharing is widely used in small sensors.
In the case of these sensors, though, each photodiode has its own microlens.
Thus, the photodiode is the pixel in these designs.

The photodiode is the light-sensitive part of a pixel. A standard bayer pixel is comprised of a photodiode, at least one microlens layer (sometimes two), and a color filter, as well as the row/column activate wiring, amplifier, and readout transistors.

In a DPAF pixel, the photodiode has been split in half, with insulating material between the two halves. Each halve has independent readout. The photodiode, despite being split, still exists below the color filter and microlenses. Therefor, there is still ONE pixel...with two photodiodes. Canon did not increase the pixel count...they increased the photodiode count.

In short, depending on the implementation, a photodiode and a pixel could mean the same thing.

Your interpretation is wrong. ;P Sorry. Go read the darn patents, and stop making assumptions.

Canon's 'dual-pixel' tech is assumed to be based on a shared-transistor design.
That is, it is a multi-photodiode design.
But since in a shared-transistor design photodiodes are effectively equivalent to pixels (as explained),
Canon's tech could be called multi-pixel design as well.

But photodiodes and pixels are not effectively equivalent. A pixel is more complex than a photodiode. A photodiode is simply a PART of a pixel. Your conflating the two for the sake of your argument, but that does not mean your conflation is valid.

So, you can stop correcting people who use dual/quad-pixel terminology, as these could in fact be used interchangeably.
The line between between a pixel and a photodiode is blurred in shared-pixel designs.
And the fact that the two photodiodes are read independently for auto-focus further
indicates that these could very well be independent pixels - if they didn't share the
same microlens and color filter.

You misunderstand shared-pixel designs. Shared pixels do not share the photodiode. Each pixel still has it's own independent photodiode. What's shared in a shared-pixel design is the readout logic...transistors. Usually, the sharing is diagonal, although some prototypical designs share directly neighboring pixels. Green pixels usually share their readout logic diagonally. Those two green pixels, however, each still have their OWN photodiode. The purpose of a shared pixel design is not to share the light-sensitive charge collector...that would be useless, since it would share each pixel's charge in one bucket, meaning you couldn't actually read them out independently.

The purpose of a shared pixel design is to save die space FOR the photodiode by reusing transistors and wiring for more than one pixel. The use of shared transistors to activate, amplify, and read the pixel has nothing to do with blurring the line between pixel and photodiode. The pixel is a vertical stack of layers of silicon materials. The photodiode is (usually) at the bottom of a physical well...it's the bit of silicon that is actually sensitive to light and converts some ratio of incident photons to free electrons (charge). Above that is a layer of translucent silicon material, usually silicon dioxide. Above that is often a microlens, and above that is a color filter array. There is sometimes buffer materials in between these layers, on top of which you finally have the primary microlens. THAT is a "pixel". The photodiode is just one part of the whole pixel. If you split the photodiode underneath all those other layers...you still have just one pixel. You have a pixel that is now capable of detecting phase, but it's still just one pixel, not two pixels. Regardless of what kind of readout logic it has...a pixel is a pixel, independent and atomic, and a photodiode is just a part of a pixel.

Also, your claim that there are exactly TWO PHOTODIODES (and that's it!) is not based on fact.
We don't know for sure if Canon's design is a dual-pixel design (your assumption) or a quad-pixel design
(my assumption).

Your assuming I am assuming. Your assumption is, once again, wrong. You are also assuming that "we" don't know anything "for sure" about Canon's sensor designs. Sorry, but again, your assumption there is WRONG. Canon has filed patents for all of their DPAF designs. Those patents are the basis for their technology...the technology that actually exists in the 70D, for example. I am not assuming. My assertions are based on actual fact as clearly and definitively defined by Canon engineers themselves.

You can go look up these patents for yourself. They aren't hard to find. Many of them have been posted right here on CR in the past. This stuff isn't some mysterious, mystical, magical sensor technology that Canon is keeping obfuscated. Obfuscation and secrecy is the worst form of protection for technology. By filing and receiving patents, Canon LEGALLY protects their work from theft by other manufacturers...they have no reason to hide or obfuscate anything.

Canon's marking is selling it as a 'dual-pixel' tech likely because it's easier this way to communicate
the concept to the general public.
But we don't know for a fact what the actual implementation is.

We DO know what the actual implementation is. Not only that, we know EXACTLY what it is. See my prior comment.

So, your TWO PHOTODIODES claim is based on marketing materials, really.
If I were you, I wouldn't put too much weight into these 8).

Again, wild assumption, and a wrong one. You assume WAY too much. You might want to verify your facts first, before putting yourself out like that. I have never based anything I've said about Canon sensor technology on marketing materials. I read patents, of which there are many thousands filed by Canon every year, and many thousands more filed by all the other entities involved in sensor research and design. I know EXACTLY what I am talking about, and it's based on actual sensor designs that have either been manufactured for commercial use, or have been prototyped and thouroughly demonstrated at one of the numerous ICS conferences around the world every year.

The only person who puts weight into something they shouldn't is you...putting a lot of weight into the validity of your assumptions.

My assumption for a quad-pixel design is based on simple geometry.
If there are just two photodiodes per pixel, these photodiodes need to be rectangular.
This would be uncommon - if not even a first in the industry.
But with a quad-pixel design, the photodiodes are square just like in any other sensor.

Considering the potential future advantages of a quad-pixel design (e.g. for a non-Bayer sensor),
I'd speculate that Canon would have invested in a quad-pixel design from the start - rather than
designing rectangular photodiodes that later would need to be made square anyway.

Just a speculation, of course - but based on some informed assumptions.

The photodiodes ARE rectangular! That's EXACTLY what they are! That's exactly how they are described in Canon's patents on the technology! : It's not a first in the industry...for decades, there have been sensors with non-square photodiodes, even non-square pixels. There have been hexagonal pixels (Fuji first released sensors with hexagonally shaped pixels with extra small "white" pixels filling in the diagonal spaces between them many years ago), triangular pixels (Sony has a prototype 50mp sensor with triangular pixels), even pixels with non-uniform pixel sizes and layouts (some sensor designs, usually from Fuji, have had large regtangular white pixels, along with a non-standard layout of smaller rectangular red, green, and blue pixels). I currently use a CCD camera for guiding my astrophotography that uses rectangular pixels, due to the use of an anti-bloom gate. Again...your making some wild assumptions that have absolutely no basis in fact. Your assumptions are FAR from informed, as well. I don't know where you think your "informing" yourself, but you really need to go right to the source...patents. You seem to think that all this technology is kept secret and obfuscated and hidden away within the bowls of "Canon the over-protective corporation". That is, once again, an assumption. Canon has decades of sensor technology filed legally as patents in countries around the world. Those patents are fully available, in complete detail, with abstracts, technical diagrams, and full-blown conceptual and functional dissection and breakdown, for review by anyone who wishes to spend the time looking them up. If patents weren't freely available, then they would be useless. Competitors have to be able to investigate what technology their competitors have already invented and patented, so they don't try inventing the same exact thing to patent themselves...that would be a patent violation. Potential licensees of patented technology need to know how the technology is implemented, so they may implement it themselves in their own products, with the added requirement of a royalty fee.

This technology is WELL KNOWN, because it has to be. "We" know EXACTLY how DPAF is designed...and it is not quad-pixel. It's, quite literally, dual-photodiode. There are now multiple patents that PROVE that FACT.

 

[the blackfox]

having fallen foul of nikons removal of the AA filter as in the d7100 and after having two of those cameras literally go tits up on me after 3000 actuations each,and hence speeding up my return to canon gear ,i hope thats not what canon are doing .i also hope that whatever this breakthrough is its going to be backward compatible with current lenses and lens technology .as in most cases when a manufacturer brings out something thats a breakthrough it usually ends up the only thing it breaks is the bank .

lots of speculation on here i can't wait to see what actually comes forth .

 

[x-vision]

I know EXACTLY what I am talking about ...

Hmm. Doesn't look like it. Let's see.

First you say this:

The photodiode is the light-sensitive part of a pixel. A standard bayer pixel is comprised of a photodiode, at least one microlens layer (sometimes two), and a color filter, as well as the row/column activate wiring, amplifier, and readout transistors.

And then you say:

In a DPAF pixel, the photodiode has been split in half, with insulating material between the two halves. Each halve has independent readout.

So, a pixel has a photodiode and readout transistors (in addition to the other stuff).
Simialrly, a DPAF pixel is a (half) photodiode ... with an independent readout.

Do you even realize that by splitting the photodiode in two, and by providing independent readouts for each half,
you have essentially created two pixels ?

I don't think you do!
That's why you can't grasp that a split photodiode is in fact a separate pixel - etched on the silicon wafer.

The microlenses and color filters are secondary - put on top of the already etched wafer.
You can put a microlense and a color filter on top of multiple pixels.
That's what Canon is doing - and what you call a pixel with a 'split photodiode'.
What you don't grasp, obviously, is that a split photodiode with independent readouts ... is two separate pixels.

The photodiode, despite being split, still exists below the color filter and microlenses. Therefor, there is still ONE pixel...with two photodiodes.

There you go. That's the part that is escaping you.
It's ONE pixel in the image, not on the silicon wafer.

There is a reason I mentioned the 'classic' 3T pixel and the shared transistor designs.
And that is to illustrate the point that a 'pixel' can be implemented in different ways.
The important distinction is that the wafer is etched in a way so that you can read photodiode charges independently.

That's what a pixel is on the wafer level.
You can certainly combine the output of multiple pixels into one.
Or put a single microlens on top of multiple pixels.
But as long as you have a photodiode and independent readout circuitry, regardless of configuration,
you have a pixel - and that part is definitely escaping you.

And the reason is that you are not technical enough to grasp the underlying principle here.
So, no, you don't know what you are talking about.

You misunderstand shared-pixel designs. Shared pixels do not share the photodiode. Each pixel still has it's own independent photodiode.

You mean just like a split photodiode with two independent readouts???

The photodiodes ARE rectangular! That's EXACTLY what they are! That's exactly how they are described in Canon's patents on the technology! :

O-o-kay. Care to share a link to least one of these patents. That should settle it, right?
So, let's settle it by you providing a link to at least one of these patents.
You have the link handy, don't you?

Look, I suggest that you drop the 'I'm the authority' attitude - because you are not an authority.
In fact, it's very clear that you don't even come from a technical background.
So, drop the attitude and let's have a friendly discussion.
That's the reason why we are all here, no?. What's with all that bullying ??

 

[jrista]

I know EXACTLY what I am talking about ...

Hmm. Doesn't look like it. Let's see.

First you say this:

The photodiode is the light-sensitive part of a pixel. A standard bayer pixel is comprised of a photodiode, at least one microlens layer (sometimes two), and a color filter, as well as the row/column activate wiring, amplifier, and readout transistors.

And then you say:

In a DPAF pixel, the photodiode has been split in half, with insulating material between the two halves. Each halve has independent readout.

So, a pixel has a photodiode and readout transistors (in addition to the other stuff).
Simialrly, a DPAF pixel is a (half) photodiode ... with an independent readout.

Do you even realize that by splitting the photodiode in two, and by providing independent readouts for each half,
you have essentially created two pixels ?

They aren't two pixels. It's two photodiodes in a SINGLE pixel. Your trying to reduce a pixel to just the photodiode. That's incorrect. Just because they have independent readout does not make them separate pixels. Both halves are still one color. There is no useful purpose to reading each half out independently for an image read. If you read a red DPAF pixel out as independent halves, you have two red rectangular results...but those results have no meaning independently. They are just two red values with half the light and twice the read noise than what you would get if you binned the two halves electronically at readout.

I don't think you do!
That's why you can't grasp that a split photodiode is in fact a separate pixel - etched on the silicon wafer.

The microlenses and color filters are secondary - put on top of the already etched wafer.
You can put a microlenses and a color filter on top of multiple pixels.
That's what Canon is doing - and what you call a pixel with a 'split photodiode'.
What you don't grasp, obviously, is that a split photodiode with independent readouts ... is two separate pixels.

If we reduce everything to the most simple sensor, monochrome, with nothing but light-sensitive silicon patches and their companion readout transitors...then you would be correct. A photodiode would then be equivalent to a pixel.

Were not talking about a simple monochrome sensor. We are talking about a bayer sensor. A sensor that has red, green, and blue PIXELS that are interpolated during demosaicing to produce full-color RGB output pixels. Each of these pixels in a bayer sensor is, at the VERY LEAST, comprised of a color filter layered on top of a photodiode. If you split the photodiode...you now have two photodiodes that read from below the same color filter. From an interpolation standpoint...you still have to combine those two halves...either electronically or digitally, to perform demosaicing.

If you completely ignore the fact that the RAW sensor readout in a bayer sensor needs to be demosaiced, then sure...you theoretically have the potential to produce two outputs per color filter. But what does that mean? How is that useful? Spatially, nothing has really changed. Whether you have two half reds, four half greens, and two half blues...SPATIALLY, they are still IDENTICAL to one red, two green, and one blue. SPATIALLY, you've gained nothing. It doesn't matter if you can read them out independently.

Your trying to imply that somehow, this increases your resolution. It does not. Just because two (or more) photodiodes underneath a single color filter can be read out independently does not change the fact that they are all the same color, and they all define the same spatial frequency as a single photodiode under that same filter. The only way those independent photodiodes could actually become useful is if you actually built microlenses to focus a cone of light onto each one independently. THEN you might actually increase luminance resolution, and you might actually have an increase in spatial resolution.

But Canon's patents do not describe a pixel structure wherein multiple microlenses are used to focus a cone of light onto each part of the split photodiode. On the contrary, the patent's CAN'T describe such a pixel structure...as then you wouldn't actually be able to perform AF functions with it. The point of DPAF is to add PHASE DETECTION capabilities to a sensor...not increase it's resolution.

The photodiode, despite being split, still exists below the color filter and microlenses. Therefor, there is still ONE pixel...with two photodiodes.

There you go. That's the part that is escaping you.

Read the above.

You misunderstand shared-pixel designs. Shared pixels do not share the photodiode. Each pixel still has it's own independent photodiode.

You mean just like a split photodiode with two independent readouts???

That is the exact OPPOSITE of a shared pixel. Shared pixels SHARE readouts. Canon's DPAF use INDEPENDENT readouts. They HAVE to use independent readouts, because they are in the same row. You cannot share pixel transistors in the same row, since columns of pixels are read out row-by-row. DPAF is essentially the opposite of a shared pixel sensor design...instead of reducing logic space and sharing logic among larger photodiodes, DPAF increases logic space and isolates logic among smaller photodiodes.

The photodiodes ARE rectangular! That's EXACTLY what they are! That's exactly how they are described in Canon's patents on the technology! :

O-o-kay. Care to share a link to least one of these patents. That should settle it, right?
So, let's settle it by you providing a link to at least one of these patents.
You have the link handy, don't you?

Look, I suggest that you drop the 'I'm the authority' attitude - because you are not an authority.
In fact, it's very clear that you don't even come from a technical background.
So, drop the attitude and let's have a friendly discussion.
That's the reason why are all here, no?. What's with all that bullying ??

I have the PDFs for those patents saved on my hard drive. I'm not going to do the legwork for you AGAIN to find the source for those downloads. I shared several links to DPAF patents the LAST time we had this debate. You clearly ignored them. If you want to educate yourself, educate yourself. You can start here (they have the patent number...go dig through the bowels of the internet on your own time to find it):

http://thenewcamera.com/canon-patent-more-sensetive-dual-pixel-cmos-af/

As for the rest, your making a LOT of assumptions, and piling assumption on top of assumption, then making bold claims about how you've discovered Canon has QPAF technology, all based on nothing but assumption and a misunderstanding of a tiny little image from a single page of ChipWorks site. If Canon had QPAF technology, they would NOT be keeping it a secret. That would be insane for them, what with the perception in the community at large being that Canon is behind on sensor tech. QPAF would be huge news for Canon. I've spent years on this forum debunking hair-brained theories like that because all it does is mislead people, give people false hopes, and otherwise confuse the issue about what any given technological advancement REALLY offers, what it REALLY allows, and how it is REALISTICALLY likely to evolve into the future. Your radically confusing the issue about DPAF. It's a very simple technology, designed for a SINGLE purpose, and it serves THAT purpose extremely well. Your trying to inflate it into something it isn't even remotely intended to be. Sorry, but I've never liked it when people make wild uninformed assumptions then boldly claim they know what their talking about. It's just something that irks me.

 

[x-vision]

That is the exact OPPOSITE of a shared pixel. Shared pixels SHARE readouts. Canon's DPAF use INDEPENDENT readouts.

Heh. You share readout circuitry between photodiodes ... to read their output independently.
What a paradox. And yet, that's exactly what the industry has been doing for a decade now (or more?).
Fascinating stuff. LOL.