Sigma DP Quattro

[Hulyss]

Some explanations from da Boss :

http://youtu.be/dZjeiE8f32Y

[eronald]

I guess the top "blue" layer is in fact (mo)nochromatic, with decent sensitivity into the greens, so the Quattros will by default be usable as a 20MP orthochromatic cameras. Anyone tried this with the Merrills?  

A well-chosen red or yellow filter might make it single layer panchromatic.

Edmund

[RobertJ]

Anyone speak Japanese?  Towards the end of the video, when he talks about the lenses, does he say whether the 19mm is the same, or updated?

[eronald]

Anyone speak Japanese?  Towards the end of the video, when he talks about the lenses, does he say whether the 19mm is the same, or updated?

I didn't catch that; my japanese is not good enough. However I think I heard some comments about how having a fixed lens ensures that there are no precision losses to camera mount and lens mount tolerances, and enable precise factory calibration. I don't think this was all marketing.

Edmund

[BernardLanguillier]

Anyone speak Japanese?  Towards the end of the video, when he talks about the lenses, does he say whether the 19mm is the same, or updated?

I do.

I have only watched the 10 last minutes or so, here is what the guy says (he is pretty funny):
- around 37:00, he explains that the spatial resolution of the quattro is similar to that of an unspecified 36mp camera and that the lack of moire in the quattro image compared to the DSLR is a clear proof that Sigma's image processing is able to compute real colour information at each pixel from the Quattro sensor. His point is that there is no loss of colour quality compared to the X3 sensor of the DPx Merrill,
- the lenses used in the Quattro DP2 and DP3 are the same as those of the DP2m and the DP3m, but the lens of the DP1 quattro was indeed re-designed compared to that of the DP1 Merrill,
- the design of the body was thought out with the intent of holding the camera with the left hand under the lens so as to maximise hand held stability and optimize image quality,
- the slightly larger size of the bodies results from the amount of high performance/memory chips embedded, said to be at the same level as the Canon 1Dx,
- the alignment of the lens and sensor is tuned/verified on the assembly line for each and every unit which ensures an image quality hard to match with interchangeable lens cameras such as DSLRs,
- the usage of a lens shutter also contributes to lower vibrations and less image quality loss compared to the plane shutter equipped DSLRs.

Overall, although they don't spell it out per se, they seem confident that the real world image quality of the Quattro series is basically currently un-matched.

Cheers,
Bernard

[ErikKaffehr]

Hi,

Quite a few good points, it makes a lot of sense.

Personally, I still feel that bayer with smaller pixels and proper OLP filtering makes more sense, but it is nice that Sigma develop their designs so consistently.

I certainly feel that Sigma is a company to watch, especially now, when mirrorless designs are getting more and more important.

Best regards
Erik

I do.

I have only watched the 10 last minutes or so, here is what the guy says (he is pretty funny):
- around 37:00, he explains that the spatial resolution of the quattro is similar to that of an unspecified 36mp camera and that the lack of moire in the quattro image compared to the DSLR is a clear proof that Sigma's image processing is able to compute real colour information at each pixel from the Quattro sensor. His point is that there is no loss of colour quality compared to the X3 sensor of the DPx Merrill,
- the lenses used in the Quattro DP2 and DP3 are the same as those of the DP2m and the DP3m, but the lens of the DP1 quattro was indeed re-designed compared to that of the DP1 Merrill,
- the design of the body was thought out with the intent of holding the camera with the left hand under the lens so as to maximise hand held stability and optimize image quality,
- the slightly larger size of the bodies results from the amount of high performance/memory chips embedded, said to be at the same level as the Canon 1Dx,
- the alignment of the lens and sensor is tuned/verified on the assembly line for each and every unit which ensures an image quality hard to match with interchangeable lens cameras such as DSLRs,
- the usage of a lens shutter also contributes to lower vibrations and less image quality loss compared to the plane shutter equipped DSLRs.

Overall, although they don't spell it out per se, they seem confident that the real world image quality of the Quattro series is basically currently un-matched.

Cheers,
Bernard

[BernardLanguillier]

Personally, I still feel that bayer with smaller pixels and proper OLP filtering makes more sense, but it is nice that Sigma develop their designs so consistently.

Erik,

I am not sure I agree. At lower ISOs, the pixel quality of the DPx Merrill is simply unmatched and viewing those files at 100% is a real pleasure. I have never seen such a pixel quality in files from other cameras, be it the D800, MFDBs or Betterlight. The only thing coming close may be the MS backs but most MF lenses I have seen are no match for the lenses of the Merrills.

Since I currently only own a DP2 Merrill, a DP3 Quattro seems a logical addition and may replace my D800 as stitching camera when bulk is an important criteria.

Cheers,
Bernard

[LKaven]

I do.

Ii yo Bernard! 

I think it's really time to mind the difference between sensels and pixels.  We often forget that bayer sensels are not pixels.  But it is easy to forget when the bayer matrix bears a superficial similarity to a video raster.  The Quattro sensor does not allow us to forget so easily. 

There was myriad ways to exploit information inherent in a complex sensor topology like the Quattro.  Very little is generally known about what the information content inherent in this sensor really is.  I feel pretty sure that there are a number of clever tricks being applied.  I can't wait until a technical paper becomes available.

[ErikKaffehr]

Hi,

I bet it will be a long wait…

I would say that is one weakness of the Foveon concept. It needs different processing from bayer, and it is not widely supported by leading raw processing tools.

But, Sigma could cooperate with some nice firm making a nice workflow tool.

Best regards
Erik

Ii yo Bernard! 

 I can't wait until a technical paper becomes available.

[Bart_van_der_Wolf]

I have only watched the 10 last minutes or so, here is what the guy says (he is pretty funny):
- around 37:00, he explains that the spatial resolution of the quattro is similar to that of an unspecified 36mp camera and that the lack of moire in the quattro image compared to the DSLR is a clear proof that Sigma's image processing is able to compute real colour information at each pixel from the Quattro sensor.

Hi Bernhard,

And it clearly shows that they do not make a proper distinction between aliasing, which the Foveon sensors inevitably create (just like any other discrete sampling sensor without OLPF), and False Color Artifacts (due to different sampling densities and positions between Green and Red/Blue in Bayer CFA designs). This improper understanding/suggestion leads to many wrong conclusions, such as lack of aliasing (where it demonstrably exists, exactly where the Nyquist frequency is exceeded, just like the theory describes).

His point is that there is no loss of colour quality compared to the X3 sensor of the DPx Merrill,

Which is not entirely true either. It's just that the Merrill middle and bottom layers are more blurry than the top layer, so the loss of color resolution is not that large in the Quattro design. But in order to avoid false color artifacting, that requires to use the 4 binned color clues from the top layer in the Quattro design, which reduces the color resolution of that layer. In other words, there will be some loss of color resolution, but not much.

Luminance resolution can be sampled mostly from the top layer, and in the Raw converter that luminance component can be used to improve the apparent resolution of the middle and bottom layer of the Quattro design. Pretty clever trade offs IMHO. However, it samples luminance at 5424 x 3616 sensel positions, and therefore 19.6 MP is the native output size. The reduced sampling density of the  middle and bottom layer of the Quattro design will result in a better Raw file storage size, which is helpful.

Overall, although they don't spell it out per se, they seem confident that the real world image quality of the Quattro series is basically currently un-matched.

It will probably be pretty good, although it will require a very elaborate Raw conversion to create a good color image from an almost monochrome capture. It will be so different from a general Bayer Demosaicing, that there will be not many other Raw converter suppliers willing to invest the amount of time required to do a good job. It therefore seems that a separate Conversion/Processing application workflow will be inevitable.

Cheers,
Bart

[Bart_van_der_Wolf]

Since I currently only own a DP2 Merrill, a DP3 Quattro seems a logical addition and may replace my D800 as stitching camera when bulk is an important criteria.

Indeed, the higher magnification factor of the 50mm lens (with reasonable DOF), and the 19.6 MP resolution in a relatively small package, are an interesting platform for a compact stitching setup.

Cheers,
Bart

[LKaven]

However, it samples luminance at 5424 x 3616 sensel positions, and therefore 19.6 MP is the native output size.

There's a bit of a leap there from sensels to pixels.  

When we're talking either about the Bayer sensor or the Foveon, sensels do not equal pixels.  All sensors interpolate, exploiting natural information, and some natural tendencies, in various ways, to render (tricolor) pixels.  What we /are/ talking about is exploiting information to make images that are "realistic", and not so much "real".  

Somewhere along the way, it pays to remember that this is not a real Nyquist domain.  This is not a discrete time-sampled domain.  There is no /true/ wave reconstruction going on in the same way that audio samples can be used to /uniquely/ determine the original waveform.  

It might help us to understand just how the first generation of Merrill sensors, at somewhere around 15M luminance-sensing positions, somehow manages to yield so much more /perceivable/ detail than the number of sensels might lead us to naively assume?

[RobertJ]

Wow, thanks Bernard.  It's like I knew what he was saying, while I don't speak a word of Japanese.

After I see how this new sensor performs, I just might buy all three.

[Bart_van_der_Wolf]

There's a bit of a leap there from sensels to pixels.

Hi Luke,

I said sensel positions. It's the sampling positions that determine the sampling density in 2D space (the focus plane). Foveon complicated matters by stacking (3) 'photosites' with (1) sampling position, but the positions are what counts for resolution. Measuring a signal 3x at the same position (the Foveon claim to faim), delivers the same spatial resolution, although probably with (slightly) different signal levels (at least caused by photon shot noise), and potentially higher color resolution (depending on sampling pitch and sampling aperture/microlens size and shape).

When we're talking either about the Bayer sensor or the Foveon, sensels do not equal pixels.

But they do, albeit different quality (broadband/narrowband color) pixels. The ISO standards also make that distinction between sampling/input position (sensor element), and output (monochrome, or RGB, or HSL, or ...) pixel. Sensor element -> Picture Element (sensel->pixel). It's really simple and straightforward.

It was complicated by some convoluted Foveon speak, because they needed a marketing tool to close the gap at that time between the Bayer CFA sensel count and the Foveon sensel count. Instead of explaining what the real differences are between single color sampling and multiple color sampling (at the same sampling position), and the differences between OLPF filtered and non-filtered aliasing (but aliasing in both(!) cases), they created more confusion than is helpful. And the confusion continues, as this thread and many others demonstrates.

Somewhere along the way, it pays to remember that this is not a real Nyquist domain.  This is not a discrete time-sampled domain.  There is no /true/ wave reconstruction going on in the same way that audio samples can be used to /uniquely/ determine the original waveform.

It is exactly the same, only the time domain is changed in a spatial (position) domain. Nyquist is as relevant in the discrete spatial domain as it is in discrete time domain. It's the basis of Digital signal processing (read all about it). Analog signals are converted into Digital signals, and digital signals are discrete signals both in amplitude and sequencing (be it in time/frequency/space or other domains).

It might help us to understand just how the first generation of Merrill sensors, at somewhere around 15M sensels, somehow manages to yield so much more /perceivable/ detail than the number of sensels might lead us to naively assume?

Not really necessary, because it is (and always has been) merely caused by the absence of an optical low pass filter (OLPF), and three color samples at the same position. There is no magic, it is straightforward DSP. It's exactly why some Bayer Sensor designs use 4x multi-sampling, and why they eliminate an OLPF. Nothing new under the sun, really, just another set of drawbacks (constant lighting requirements between piezo-element shifted sub-exposures, and more image magnification to reduce aliasing tendency).

It's IMHO also time to drop the mumbo jumbo (it only leads to impossible to explain assumptions and no real solutions/explanations), and just appreciate the benefits of the various imperfect but very usable methods of image capture. If only battery technology progressed faster, because that would allow more elaborate in camera processing with a decent battery life.

Cheers,
Bart

[LKaven]

Hi Bart,

We are both saying a number of true things, so it's worth doing a quick stock take.

We both agree that Nyquist theory can be, and is, fruitfully applied to spatial sampling problems.

However, there are some differences.  There is no continuous function being reconstructed in digital photography.  The positioning of sensels has only a contingent relationship to pixel values in the output raster, not a necessary one!

I maintain the there is an important distinction between sensels and pixels.  Pixels are positionally-coordinated tristimulus values.  Sensels are abstract detectors.  Nothing a priori necessitates a 1-to-1 mapping of positional sensels with raster positions.  

In my view, there is nothing to suggest a determinate answer to "what is the resolution of the Foveon sensor, in /pixels/?"  And I believe that to define the resolution of the Foveon sensor as the number of top-layer sensels is a naive view.  As with /all/ digital image sensors, there is a certain amount of confabulation.

By my observations, the lack of an AA filter does not explain the differences between a 24MP APS-C sensor without an AA filter, and the output of a DPxM.  Where do the extra locations in a super-large rendering come from?

[Bart_van_der_Wolf]

Hi Bart,

We are both saying a number of true things, so it's worth doing a quick stock take.

We both agree that Nyquist theory can be, and is, fruitfully applied to spatial sampling problems.

However, there are some differences.  There is no continuous function being reconstructed in digital photography.  The positioning of sensels has only a contingent relationship to pixel values in the output raster, not a necessary one!

Hi Luke,

I don't understand what you are saying here. Digital signals are discrete in both amplitude and in sequence, it's the definition of a digital signal. The signal at the sampling position is quantized by the ADC and becomes digital. The output pixel is also discrete in both amplitude and position/sequence. Sensel position becomes pixel position, although signal processing will alter the amplitude value (amongst others by White balancing, Gamma precompensation, and in case of Bayer CFA samples by demosaicing, i.e. reconstruction of the missing information which is a bit less accurate than a direct sample, and in case of a Foveon sensor by channel separation).

I maintain the there is an important distinction between sensels and pixels.  Pixels are positionally-coordinated tristimulus values.  Sensels are abstract detectors.  Nothing a priori necessitates a 1-to-1 mapping of positional sensels with raster positions.

I'm not sure what you are suggesting here, but one sampling position leads to one pixel. If not, then one is binning or resampling, but both affect the original sampling density and therefore resolution and its implicit Nyquist frequency. Again, the amplitude will change from a quantized electron count (an Analog Digital Unit or ADU) to an RGB coordinate. Input sensel and Output pixel are related by a conversion process that changes the amplitude value into something that has a visual meaning, a grayscale level or color.
  

In my view, there is nothing to suggest a determinate answer to "what is the resolution of the Foveon sensor, in /pixels/?"  And I believe that to define the resolution of the Foveon sensor as the number of top-layer sensels is a naive view.  As with /all/ digital image sensors, there is a certain amount of confabulation.

Not really, just take a shot of an appropriate (free) test chart, and the resolution will be shown, including the tendency for producing aliasing artifacts. The is hardly any ambiguity possible, an a reasonably hard number can be stuck to the resolution limit (although there is more to that story than the limit alone).

By my observations, the lack of an AA filter does not explain the differences between a 24MP APS-C sensor without an AA filter, and the output of a DPxM.  Where do the extra locations in a super-large rendering come from?

What extra locations? Aliasing artifacts?

Cheers,
Bart

[capital]

...

Which is not entirely true either. It's just that the Merrill middle and bottom layers are more blurry than the top layer, so the loss of color resolution is not that large in the Quattro design. But in order to avoid false color artifacting, that requires to use the 4 binned color clues from the top layer in the Quattro design, which reduces the color resolution of that layer. In other words, there will be some loss of color resolution, but not much.

...

Hi Bart,
Where is the information about "blurring" of middle and bottom layers coming from for Merrill type x3 sensors?

As far as I can tell, the reason they are doing a 1:1:4 layout is borne out in the Quattro Patent information disclosures which indicate that while the preferred sampling modality is a 1:4:1 stack the needed wiring is too bulky, which ends up interfering with light collection and noisey signals. And that a 1:1:4 is "good enough" for now as they state "although the blue channel's spectral response is less ideal (farther from a theoretically ideal luminance spectral sensitivity curve) than is the green channel's spectral response, because the full resolution blue layer of each such implementation responds to green and red light as well as blue light. These implementations of the invention can adequately capture high resolution luminance information, while their full resolution readout of the top (blue) layer and lower resolution readout of the other (green and red) layers also provides advantages (e.g., compactness, noise improvement, and reduction in the number of contacts that must be provided to sensors in the green and red layers) that cannot be realized by full resolution readout of the intermediate (green) layer of a "1-4-1" array and lower resolution readout of the blue and red layers of the "1-4-1" array."

[Bart_van_der_Wolf]

Hi Bart,
Where is the information about "blurring" of middle and bottom layers coming from for Merrill type x3 sensors?

Hi,

It's from observation (check the largely monochrome individual 'channels' as obtainable from a simple DCRaw dissection of pure Raw channel data yourself), caused by scattering and defocus, and from one of the inventors of the Foveon sensor, Richard Lyon, himself (after he sold his interests in the technology he can speak more freely) here (although he might be referring to the new Quattro design in particular).

The large area of the lower-level pixels is the ideal anti-aliasing filter for those layers; the top layer is not compromised by the extra spatial blurring in the lower layers, so it provides the extra high frequencies needed to make a full-res image.

His comments here are also notable.

Anyway, the top layer is not white, not luminance, not blue, but a sort of panchromatic blueish that turns out to work well enough for getting a high-frequency luminance signal.  We did a lot of experiments and amazed ourselves how "well enough" the 1:1:4 worked; it was not obviously going to be a good thing, but turned out awesome.

Unfortunately, which is why I hesitated to supply the link, even Dick Lyon (an electrical engineer) gets the DSP aspects of the Bayer CFA (for which Bryce Bayer deservedly received a Nobel prize) wrong (Bayer CFAs do not sample 25% of a pixel sampling aperture at best). His concept apparently involves 4 Bayer CFA input sensels per 1 output pixel (which is totally not what happens upon capture and subsequent Bayer CFA demosaicing).

Cheers,
Bart

[LKaven]

Hi Bart,

I think you are not seeing the thrust of my argument.  Or perhaps I am giving you an argument that you readily concede, and vice-versa.  I'm not sure.  Either way, I understand your temptation to remind me of fundamentals here, but these are things I've been acquainted with for many years. 

I am responding in large part to the mistaken tendency to identify sensels with pixels.  I am also making note of the sense in which digital photography uses confabulation in signal reconstruction, and that this means that the number of final "pixels" is variable with respect to the number and type of sensels  You seem to agree with that.

I also in fact agree with you that the positioning of the sensels is "pixel-like" and that these positional sensels do mean something in computing any reasonable measure of "resolution", e.g., /optical/ resolution.

And perhaps there is no mystery as to why the Merrill in the DPxM is sold as a "46MP" sensor, and capable of "convincingly" and "believably" producing more than the 15.3M sensel locations would suggest.  Why does channel separation work so well here?

I use the terms "convincing" and "believable" here to indicate that this is the best that we can say about confabulated information. 

And I contrast this with classical Nyquist theory, which is for continuous functions.  Compare the case of digital audio with the case of digital photography.  Digital audio begins as a continuous function in the continuous variations in signal produced by a sound transducer.  That continuous function is sampled into a discrete-time domain of course.  But in the end, what comes to your speakers is a reconstruction of the continuous function.  In ideal cases, where (i) the original signal lies entirely within the bandpass filter, and (ii) the precision is infinite, then the reconstruction should be identical to the original, as a matter of analytical fact.  There is no confabulation in such cases.  In digital photography, there is substantial confabulation.

Also, in the case of continuous Nyquist functions, there is no other information needed other than the samples themselves (along with sample rate) in order to perform an exact reconstruction.  In digital photography, implied information content is exploited to whatever extent possible in order to enhance the believability of the results. 

In the end, a pixel rating involves believability coupled with a good measure of truth.

Sensel position becomes pixel position, although signal processing will alter the amplitude value (amongst others by White balancing, Gamma precompensation, and in case of Bayer CFA samples by demosaicing, i.e. reconstruction of the missing information which is a bit less accurate than a direct sample, and in case of a Foveon sensor by channel separation).

[fdisilvestro]

And I contrast this with classical Nyquist theory, which is for continuous functions. 

I don't understand this, are you claiming that the light intensity in the spatial domain does not behave like a continuos function and Nyquist don't apply?

Compare the case of digital audio with the case of digital photography.  Digital audio begins as a continuous function in the continuous variations in signal produced by a sound transducer.  That continuous function is sampled into a discrete-time domain of course.  But in the end, what comes to your speakers is a reconstruction of the continuous function.  In ideal cases, where (i) the original signal lies entirely within the bandpass filter, and (ii) the precision is infinite, then the reconstruction should be identical to the original, as a matter of analytical fact.  There is no confabulation in such cases.  In digital photography, there is substantial confabulation.

In digital imaging you have intensity of light and frequency. Since you are sampling only intensity you need another means to reconstruct frequency, which is necessary to reconstruct color. In other words, digital sensors "count" photons but do not record the wavelenght or energy associated to those photons, so an indirect method has to be applied to reconstruct the color information.

Also, in the case of continuous Nyquist functions, there is no other information needed other than the samples themselves (along with sample rate) in order to perform an exact reconstruction.  In digital photography, implied information content is exploited to whatever extent possible in order to enhance the believability of the results. 

If you had enough multi-spectral samples you would need only the sample themselves (and the sample rate) to perform an exact reconstruction