Sigma DP Quattro has 4:1:1 (four top/1 middle/1bottom photosite) design

[Kendall Helmstetter Gelner]

I thought it would be interesting to ask people here what they thought of Sigma's new layered design for the DP Quattro cameras - it has four top photo sites (roughly blue) stacked on top of a single middle and bottom layer (green and red).

So, as Sigma puts it the raw resolution is:

T:5,424×3,616

M:2,712×1,808

B:2,712×1,808

Loading…

www.sigma-global.com

Thoughts?

 

[SigmaTog]

A new era for Foveon / Sigma for sure...........

We'll just have to wait a few more sleeps or months ?

ΣigmaTog

 

[Raist3d]

I thought it would be interesting to ask people here what they thought of Sigma's new layered design for the DP Quattro cameras - it has four top photo sites (roughly blue) stacked on top of a single middle and bottom layer (green and red).

So, as Sigma puts it the raw resolution is:

T:5,424×3,616

M:2,712×1,808

B:2,712×1,808

http://www.sigma-global.com/en/cameras/dp-series/#/dp2

Thoughts?

My thoughts are (just to put them under your thread as I sort of already mentioned them elsewhere and in the messaging), that B&W wins, color will get interesting. It's no longer 3X but it should still be easier to interpolate than Bayer me thinks.

On first pass the photo sensor amount of say a Bayer 24 MP Sony sensor distributes as 6 MP red & blue, 12 MP green. So the new Foveon is about ~4.42 MP Red & Green, ~19.6 MP Blue, but spatially the red & blue of the Bayer are separate and exclusive while the Sigma 4.42 area does measure X3 in that area plus 3 extra blue measurements.

Hmmm green is something the eye is more sensitive to, so I wonder why not make the Green layer the one rich in spatial locations, but maybe it's a layer-manufacture issue and blue has to go on the top.

I thin the color shots will be ok but may have some color artifacts occasionally (and by that I don't mean necessarily major ones compared to say Bayer or Xtrans). Color moire is possible now though, but probably can be mitigated more than Bayer- possibly.

One thing vs the Merrill is that since the Merrill had so much noise particularly in the red layer, this new sensor even though maybe vs a noiseless ideal Merrill sensor would do a bit worse in some areas, in reality it may very well do better as color constancy / ISO improve.

There's also finally 14 bit raw- hopefully it uses at least true 13-bits of data - many of the original 14-bit raws were marketing as the extra 2 bits were not used, though today's sensor are good enough to finally do that. But I expect this one to at least do better than 12.

 

[hjulenissen]

Even though 2 color channels have reduced spatial resolution, they have full area coverage. This, together with the full-resolution top layer might allow for less OLPF (I assume none) while keeping aliasing in check.

I am most interested in what they gain by doing this. It is sort of an "irregular" design compared to the old one?

It shall be interesting to see what the bombastic "give me full spatial rgb resolution at each sensel or give me nothing" people will say to this...

-h

 

[Kendall Helmstetter Gelner]

Even though 2 color channels have reduced spatial resolution, they have full area coverage. This, together with the full-resolution top layer might allow for less OLPF (I assume none) while keeping aliasing in check.

I am most interested in what they gain by doing this. It is sort of an "irregular" design compared to the old one?

It is indeed irregular compared to previous designs... But I'm pretty sure I saw speculation or something on this in the past, Raist3D mentioned he had posted a similar idea some time ago.

It shall be interesting to see what the bombastic "give me full spatial rgb resolution at each sensel or give me nothing" people will say to this...

Ahh... but was it about full spatial RGB resolution, or about not discarding any input data (as the filters do). It's easy enough to transition from one idea to the other as a key property... if the results are still good.

In the end, that's the real key. How will detail come across from this sensor to the older one? If it can maintain nearly the same level of detail then people will be fine with slightly less "pure" RGB values going into each sensel. With the blue (top) sensor count increasing it could be the tradeoffs are about even and you end up with roughly equivalent to the same amount of detail while reducing noise and improving high ISO support.

 

[Raist3d]

Even though 2 color channels have reduced spatial resolution, they have full area coverage. This, together with the full-resolution top layer might allow for less OLPF (I assume none) while keeping aliasing in check.

I am most interested in what they gain by doing this.

ISO performance.

It is sort of an "irregular" design compared to the old one?

Not sure what you mean with that.

It shall be interesting to see what the bombastic "give me full spatial rgb resolution at each sensel or give me nothing" people will say to this...

Definitively true. It may not be a Bayer sensor but it sure has to resort to color interpolation now spatially, while the previous Foveons in theory did not (saying in theory due to possible noise handling strategies).

-h

 

[Raist3d]

Even though 2 color channels have reduced spatial resolution, they have full area coverage. This, together with the full-resolution top layer might allow for less OLPF (I assume none) while keeping aliasing in check.

I am most interested in what they gain by doing this. It is sort of an "irregular" design compared to the old one?

It is indeed irregular compared to previous designs... But I'm pretty sure I saw speculation or something on this in the past, Raist3D mentioned he had posted a similar idea some time ago.

It shall be interesting to see what the bombastic "give me full spatial rgb resolution at each sensel or give me nothing" people will say to this...

Ahh... but was it about full spatial RGB resolution, or about not discarding any input data (as the filters do).

There's been both claims. When you think of it you are discarding spatial resolution vs the older design for full color.

It's easy enough to transition from one idea to the other as a key property... if the results are still good.

In the end, that's the real key. How will detail come across from this sensor to the older one? If it can maintain nearly the same level of detail then people will be fine with slightly less "pure" RGB values going into each sensel. With the blue (top) sensor count increasing it could be the tradeoffs are about even and you end up with roughly equivalent to the same amount of detail while reducing noise and improving high ISO support.

I really think for B&W it will be a non issue. For color I think we are going to see new artifacts that we did not see before (color moire can happen now). But we will see less other artifacts the other design had due to noise, at least in some of the Iso ranges.

 

[DavidMillier]

I would think that this approach is going to cause a few heartaches for those who espoused Foveon on purist philosophical grounds.

I was certainly attracted to the Foveon approach because of its "digital artefact free" potential but because of the insistence on not using a OLPF AA filter Foveon never completely lived up to its potential in that respect. This new approach whilst retaining layers and silicon filtering would seem to be a move at least partway towards the flaws of the CFA approach. A trade off presumably Sigma feels needs to be made. The pragmatist in me says let's wait and see what impacts the new architecture has in real world shooting. Maybe it will turn out to be he most effective approach to using a layered architecture.

Certainly, a striking and thought provoking introduction...

 

[hjulenissen]

It is sort of an "irregular" design compared to the old one?

Not sure what you mean with that.

I only mean that it seems more complex (different planes are treated differently), meaning that there are two different systems instead of one to design, manufacture, etc.

-h

 

[D Cox]

I thought it would be interesting to ask people here what they thought of Sigma's new layered design for the DP Quattro cameras - it has four top photo sites (roughly blue) stacked on top of a single middle and bottom layer (green and red).

So, as Sigma puts it the raw resolution is:

T:5,424×3,616

M:2,712×1,808

B:2,712×1,808

http://www.sigma-global.com/en/cameras/dp-series/#/dp2

Thoughts?

My thought is that it is easier to think of the three layers in terms of Lab colour space, rather than RGB.

The top layer is L and the other two provide a and b.

For most images, as you can see by experimenting in Photoshop, the a and b channels can be much lower resolution than L with little effect. Indeed, common advice is to sharpen L and blur a and b (to reduce colour noise).

There would still be loss of resolution for fine colour detail where the colours are exactly matched in brightness, as in some textiles or perhaps a distant field with poppies.

 

[John Sheehy]

ISO performance.

That would be true if Foveon found a way to decrease read noise per unit of sensor area with biger pixels, which is not a universal given, but true of companies like Canon, who seem to be trapped into sensors that don't perform as well with higher pixel densities.

Definitively true. It may not be a Bayer sensor but it sure has to resort to color interpolation now spatially, while the previous Foveons in theory did not (saying in theory due to possible noise handling strategies).

How so? This is basically a 4.9MP Foveon sensor with 19.6MP in the "blue" channel; this can't possibly have more artifacts than a straight 4.9MP Foveon sensor would. What's "missing" in the "green" and "red" channels is resolution; not sensor coverage.

 

[DavidMillier]

Colour aliasing from mismatched channel resolution.nyquist limits?

 

[John Sheehy]

Colour aliasing from mismatched channel resolution.nyquist limits?

I don't see why. You have all the information you get from a 4.9MP Foveon sensor, with a bonus of 19.6MP from the top layer. How could that possibly result in artifacts not present in a straight 4.9MP Foveon?

The potential for tremendous color aliasing in a Bayer sensor without an AA filter is not because of the resolution of different color channels, but because of the fact that each geometrical location only witnesses a fraction of the visible light spectrum. The Quattro still witnesses nearly the entire focal plane in all three filtrations.

 

[D Cox]

ISO performance.

That would be true if Foveon found a way to decrease read noise per unit of sensor area with biger pixels, which is not a universal given, but true of companies like Canon, who seem to be trapped into sensors that don't perform as well with higher pixel densities.

Definitively true. It may not be a Bayer sensor but it sure has to resort to color interpolation now spatially, while the previous Foveons in theory did not (saying in theory due to possible noise handling strategies).

How so? This is basically a 4.9MP Foveon sensor with 19.6MP in the "blue" channel; this can't possibly have more artifacts than a straight 4.9MP Foveon sensor would. What's "missing" in the "green" and "red" channels is resolution; not sensor coverage.

I agree.

There are no gaps in the two lower channels that would have to be filled in by interpolation.

 

[Lenni Vilen]

The new sensor is fantastic innovation. As we all know, the top layer is the important one and the bottom ones just play a support role to help define perfect color. Just likey eye, no perfect color points are needed for every point, but perfect luminance or resolution. Sigma knows this so new sensor gives up a little that is not needed and improves the rest of the package. Compared to the guesswork of bayer this sensor has 80 megapixel resolution (Sigma plays nice and understates it in their words) and without the errorguesses of bayer which lead to weirdoes and color flaws. So in practise one could say it's worth more than 100 mps!!! Truly fantastic.

The new big pixels and new pixel electronics give so low noise that silky smooth is the right way to talk about this new Foveon. The dynamic range should easily beat all the full frame cameras and be at least equal to the medium format, if not better. And the ISO performance is now peerless - just imagine how much light a big pixel captures when there is no color filter throwing out most of the light! This sensor captures more than any full frame camera and does it with more style. And because no guesswork and sharpening needed, the low lit picures will be much better than anything before seen from any camera!

This camera is revolution which makes world go round to by Siigmas! Soon all the other cameras will not sale any more as people see the true colours and maximum resolution with dynamism unparalled and in nightime too. Now if Sigma comes out with a DSLR with thjis sensor and improved focusing, all the pros in the world will move and all we see in the next olumpics is Sigma lenses!

Thank you Sigma!

 

[Iliah Borg]

The new sensor is fantastic innovation.

Really? Or just the second thought?

--

About LibRaw | LibRaw

www.libraw.org

 

[hjulenissen]

I agree.

There are no gaps in the two lower channels that would have to be filled in by interpolation.

<rant>

In all likelihood, every single image you have ever printed have been interpolated. Bayer, Foveon, no matter what. If interpolation is bad, then digital images are bad.

</rant>

But there are issues with using a rectangular spatial pre-filter (i.e. area-filling sensels), in that the frequency response is quite bad, compared to a well-behaved lowpass filter.

In other words, this sensor with no pre-filtering and a sharp lens should have the potential for significant amounts of aliasing in the lowres channels.

-h

 

[D Cox]

The new sensor is fantastic innovation. As we all know, the top layer is the important one and the bottom ones just play a support role to help define perfect color. Just likey eye, no perfect color points are needed for every point, but perfect luminance or resolution. Sigma knows this so new sensor gives up a little that is not needed and improves the rest of the package. Compared to the guesswork of bayer this sensor has 80 megapixel resolution (Sigma plays nice and understates it in their words) and without the errorguesses of bayer which lead to weirdoes and color flaws. So in practise one could say it's worth more than 100 mps!!! Truly fantastic.

I would want to see real results before holding a celebration.

It does seem very likely that noise at moderately high ISO settings will be reduced. It is not clear (and may not be until there have been a couple of firmware updates) whether image quality at ISO 100 will be slightly better or not quite as good. It will probably depend on the subject matter.

It's a fair bet that the black-and-white resolution will go up slightly.

 

[D Cox]

I agree.

There are no gaps in the two lower channels that would have to be filled in by interpolation.

<rant>

In all likelihood, every single image you have ever printed have been interpolated. Bayer, Foveon, no matter what. If interpolation is bad, then digital images are bad.

</rant>

But there are issues with using a rectangular spatial pre-filter (i.e. area-filling sensels), in that the frequency response is quite bad, compared to a well-behaved lowpass filter.

In other words, this sensor with no pre-filtering and a sharp lens should have the potential for significant amounts of aliasing in the lowres channels.

But a bit less overall than in the SD15 and other Foveon cameras of that generation, because only the two layers are low res.

We will see when real life images appear.

 

[Kendall Helmstetter Gelner]

I agree.

There are no gaps in the two lower channels that would have to be filled in by interpolation.

<rant>

In all likelihood, every single image you have ever printed have been interpolated. Bayer, Foveon, no matter what. If interpolation is bad, then digital images are bad.

</rant>

They are; Foveon is the least bad option.

But there are issues with using a rectangular spatial pre-filter (i.e. area-filling sensels), in that the frequency response is quite bad, compared to a well-behaved lowpass filter.

In other words, this sensor with no pre-filtering and a sharp lens should have the potential for significant amounts of aliasing in the lowres channels.

Which will be obllviated by measured variance the top layer (which still "sees" to a lesser extent variations in the bottom layers).