Moving from 35mm to Medium Format. Need guidance (Sorry for yet another thread!)

[haplo602]

I do not like the Hasselblad strategy of a lock in either (With phaseOne, I could get an AFD III as a backup body for not too much money eventually). So yes, I am more or less on board the team PhaseOne train.

Thanks in advance,
Sandeep

To my knowledge, AFD III does not support SK leaf shutter lenses, so it is not a good backup body if your main work is flash with those lenses. Basicaly there is no combination of film back and SK leaf shutter lens on Phase One that works.

I'd love to be proven wrong however.

[Phil Indeblanc]

From the get go I had little sense that you were convinced you needed a MF, and all this talk to test that notion at least gives you assurance you're on the right track.
I think it would be a great choice, and it is a different way of working that I think makes photography more pleasurable in some ways. I hope people can provide you with the samples for you to make your final pick.
I'm a product shooter myself and why I always look for the smaller sensor out of MF is the DOF :-) But I have been so happy with the results, I keep on with it.

When you decide, I would highly recommend Doug for your purchase, as you end up getting a part of Doug to take home. That's how dedicated and supportive he is. Which will prove priceless. Best of luck. I know a lot of people have learned one or two things from this thread, I know I have, and will be looking at a DP1/2/3 for my point and shoot travel camera real soon! :-)

[Iliah]

I'm simply saying thinking of Foveon (and Bayer, for that matter) in terms of colour response is tried and tired approach. Yes, spectral response was for illustration purpose only.

[eronald]

I'm simply saying thinking of Foveon (and Bayer, for that matter) in terms of colour response is tried and tired approach. Yes, spectral response was for illustration purpose only.

I was partly trying to be funny.

As *you* know that depth of penetration is the key to color separation in the Foveon, you also certainly know that a given spectrum projected on a ray will get trapped closer -frequency by frequency, and in terms of probability - to the surface if the light ray impacts at an angle, ie you may (will) see the equivalent of a red-shift for diagonal rays. So ideally a raw converter now needs to invert both the depth-exponential channel crosstalk (overlay actually), the geometry of the stacked sensor wells, the geometry of the projected rays,  including the effects of the lens f-stop. That is forgetting about the interesting effects of chromatic aberration. I wonder whether Foveon really have a nicely written up recipe for handling all of this perfectly in the raw converter? And whether Sigma provide detailed information about their lenses and the sensor which guarantee the algorithms will work?  It sounds like a horrid inverse problem - the sort of problem only an applied mathematician can love.

I believe that this is what actually stopped brands N, C etc from investing in Foveon technology - too many variables, the maths is really involved, it's not just a 14 year old let loose on a computer shifting bytes, you need accurate experimental data for each sensor, and in the end only lens-sensor codesign will give you good results, or rather a system easy to invert.  In contrast, Bayer technology is a really good fit for reusing existing lenses and a fairly simple solution to the inverse problem for Bayer gives results acceptable to most users, with little computational effort, and no real necessity for characterising lens or sensor, as  ACR and Lightoom's initial versions have demonstrated  . History has shown that Foveon may be better but Bayer is good enough. And so, I am afraid is Lightroom


Edmund

[Bart_van_der_Wolf]

I was partly trying to be funny.

As *you* know that depth of penetration is the key to color separation in the Foveon, you also certainly know that a given spectrum projected on a ray will get trapped closer -frequency by frequency, and in terms of probability - to the surface if the light ray impacts at an angle, ie you may (will) see the equivalent of a red-shift for diagonal rays.

Hi Edmund,

Exactly, and probably one of the reasons that larger dimensions of the sensor are even more problematic, unless coupled with (dedicated) telecentric lens designs.

So ideally a raw converter now needs to invert both the depth-exponential channel crosstalk (overlay actually) and the effects of the lens at the given f-stop. I wonder how you you are handling this in the raw converter?

Well, it's all a matter of processing the Raw sensor data, number crunching, but combine angle of incidence (with wider apertures potentially projecting a multitude of ray angles), combined with many more channels than just R/G/B only to derive multi-channel color with a matrix multiplication, and we're talking about a dedicated convertor that will take its sweet time to churn out one image at a time in a pace comparable with e.g. DxO's PRIME noise reduction.

Doable, but not fun when time is an issue.

Cheers,
Bart

[eronald]

Hi Edmund,

Exactly, and probably one of the reasons that larger dimensions of the sensor are even more problematic, unless coupled with (dedicated) telecentric lens designs.

Hi Bart -

I'm happy you agree with my abstract analysis.

An interesting experiment would be to take an image wide open and at F8 on the existing cameras, and see what happens to color rendering at the field edges

Edmund

[Iliah]

SD1 metadata contains experimental data for flat field (luma and chroma) for 3 focal lengths at 4 aperture values each. The rest is to be covered by interpolation and extrapolation. The shortest length is 30mm, that is nearly 50 mm equivalent. It is not very useful for extrapolation for wider lenses, say, 24 mm equivalent. Also the data is obtained from select lenses and does not work very well with other lenses. Sigma are making a mistake by not providing end user with calibration procedures. Even their service facilities can't calibrate for a lens. The algorithms they use achieve nice results for those select lenses, but that can be improved grossly if red shift factor would be taken into account. In fact, the calibration quality and algorithmic improvement between pre-Merrill metadata and Merrill is very substantial; and the Merrill sensor structure is a gigantic step forward in terms of colour consistency compared to pre-Merrill. I think there is a possibility that current price drop is to clear the stock a little to make room for some new camera models. However I'm not so sure going to larger sensor format will not bring old issues back. Even with 1.6x crop Merrill the 3D-map of the sensor field looks ugly. If I would be working on the sensor first thing to consider is to improve noise so that 14-bit for at least 2 upper layers start to make sense. Second, a flat field calibration device for at least 2 colour temperatures, available at service facilities is a necessity. Third, lens mount rigidness and sensor-to-mount parallelism tolerance need to be improved (and that is from just a month of shooting).

[eronald]

They must be on their 5th gen or so design by now - maybe they have some new structural mitigation eg. microlenses/light-pipes for the red-shift?

I guess this conversation will explain to Sigma lovers why there is a lack of widespread software support, and why new models appear slowly.

Maybe they are waiting for us to design cheap test devices.
 
Edmund

SD1 metadata contains experimental data for flat field (luma and chroma) for 3 focal lengths at 4 aperture values each. The rest is to be covered by interpolation and extrapolation. The shortest length is 30mm, that is nearly 50 mm equivalent. It is not very useful for extrapolation for wider lenses, say, 24 mm equivalent. Also the data is obtained from select lenses and does not work very well with other lenses. Sigma are making a mistake by not providing end user with calibration procedures. Even their service facilities can't calibrate for a lens. The algorithms they use achieve nice results for those select lenses, but that can be improved grossly if red shift factor would be taken into account. In fact, the calibration quality and algorithmic improvement between pre-Merrill metadata and Merrill is very substantial; and the Merrill sensor structure is a gigantic step forward in terms of colour consistency compared to pre-Merrill. I think there is a possibility that current price drop is to clear the stock a little to make room for some new camera models. However I'm not so sure going to larger sensor format will not bring old issues back. Even with 1.6x crop Merrill the 3D-map of the sensor field looks ugly. If I would be working on the sensor first thing to consider is to improve noise so that 14-bit for at least 2 upper layers start to make sense. Second, a flat field calibration device for at least 2 colour temperatures, available at service facilities is a necessity. Third, lens mount rigidness and sensor-to-mount parallelism tolerance need to be improved (and that is from just a month of shooting).

[jerome_m]

My understanding is that (correct me if I am wrong), if you want a backup body for an H3/H3II/H4/H5, the backup body must be from the same generation and has to be sent back to the factory along with the back for calibration.

I don't think so, but I am not entirely sure. I think what you are saying is true for the H3DII, H4D and H5D bodies, which are locked to a specific back (so you need to send them to be unlocked for another back). I don't think it is true for the bodies which can use a film magazine: H1, H2, H3 and H4X.

[ondebanks]

Indeed, and it also shows that there is hardly any color separation between the channel responses. There is more overlap than separation, which also shows when examining the Raw data (which almost looks like a monochrome RGB image). There is significant mathematical separation and amplification required to produce (saturated) color, which also explains the relatively poor high ISO performance of Foveon sensors. Also skin tone color in the shadows is pretty poor.
It is almost surprising how a color image can be calculated from that source data, but it also demonstrates that with clever post-processing almost any Raw data can be made more acceptable.

Thanks for the spectral response plot, Bart. Is that for the current Foveon sensors in the DP3 and SD1 camera range? Presumably their IR-blocking filters would not significantly alter the visible light response in that plot.

A few things strike me about this:
- The Total spectral response at each pixel is admirably high. This sensor collects photons like no other found in a "regular" photographic camera. So luminance noise should be very low.
- The maths involved in extracting "normal" colours, and the resulting chroma noise issues you predict, remind me of the similar issues with my old Kodak DCS 720x. That had a CMY CFA, so each pixel received two colours - the Foveon receives three, a normal RGBG Bayer receives one. Designed for high q.e. and hence high ISOs, the 720x did have less accurate colours and a tendency for sometimes strange chroma noise thanks to the maths involved in disentangling the RGB signals from the recorded CMY signals.
- B&W images should be fantastic - especially at high ISO, with all that q.e. pulling in light. B&W requiring no colour-disentangling maths and no Bayer interpolation should be really clean and sharp.
- Narrowband imaging would also be highly profitable with this camera. Normally if you put e.g. a red nebula filter over a DSLR, you only get signal in 1 out of every 4 pixels (the R in RGBG). With the Foveon, you'd get signal in every pixel, and at a high q.e. as well!

My only doubt - what are the long exposure dark current characteristics of the Foveon?

Ray

[Iliah]

Thanks for the spectral response plot, Bart. Is that for the current Foveon sensors in the DP3 and SD1 camera range? Presumably their IR-blocking filters would not significantly alter the visible light response in that plot.

http://www.avcemporium.com/foveon-f13-and-dp13-cmos-layered-color-image-sensors/

[Bart_van_der_Wolf]

Thanks for the spectral response plot, Bart. Is that for the current Foveon sensors in the DP3 and SD1 camera range? Presumably their IR-blocking filters would not significantly alter the visible light response in that plot.

Hi Ray, I repeated the plot that Iliah attached to reply #135. I remembered something similar from an older PDF on the Foveon website from the early Foveon days. I do not know if there has been fundamentally much development since Foveon sold their technology to Sigma.

A few things strike me about this:
- The Total spectral response at each pixel is admirably high. This sensor collects photons like no other found in a "regular" photographic camera. So luminance noise should be very low.

Well, it helps that there are no CFA filters to absorb part of the luminance, but the total number of photons that do get converted (in an efficient way it seems) for 3 'channels' have to be stored in roughly the same area as a single channel for a CFA filtered sensel. However, simplistically one could say that only 1/3rd of a well depth is available for storage of each channel of a Foveon chip. In practice it's more complex than that, but the limit on total recorded photons remains (roughly the same per unit area), and thus shot noise per channel remains an issue (as evidenced by the higher ISO perfomance). It also produces huge Raw files with R+G+B-ish data per pixel, and slower per image write speeds.

- The maths involved in extracting "normal" colours, and the resulting chroma noise issues you predict, remind me of the similar issues with my old Kodak DCS 720x. That had a CMY CFA, so each pixel received two colours - the Foveon receives three, a normal RGBG Bayer receives one. Designed for high q.e. and hence high ISOs, the 720x did have less accurate colours and a tendency for sometimes strange chroma noise thanks to the maths involved in disentangling the RGB signals from the recorded CMY signals.

That's correct. The C/M/Y filtration helps with capture sensitivity because the individual filters are roughly 2/3rd transparent for the visible spectrum instead of 1/3rd of the R/G/B filters, but the subtraction required to separate RGB from CMY adds noise.

- B&W images should be fantastic - especially at high ISO, with all that q.e. pulling in light. B&W requiring no colour-disentangling maths and no Bayer interpolation should be really clean and sharp.

Yes, they produce good Black and White images, but aliased due to the absent OLPF. With a small enough sensel pitch and a large enough sensor array that would not necessarily be a major hurdle, and diffraction kind of helps to reduce some of it.

- Narrowband imaging would also be highly profitable with this camera. Normally if you put e.g. a red nebula filter over a DSLR, you only get signal in 1 out of every 4 pixels (the R in RGBG). With the Foveon, you'd get signal in every pixel, and at a high q.e. as well!

Yep.

My only doubt - what are the long exposure dark current characteristics of the Foveon?

I don't recall specifics, but the complexity of the production process may create some issues which surface when pushed into warmer operating temperatures and longer integration times.

Cheers,
Bart

[synn]

I don't think so, but I am not entirely sure. I think what you are saying is true for the H3DII, H4D and H5D bodies, which are locked to a specific back (so you need to send them to be unlocked for another back). I don't think it is true for the bodies which can use a film magazine: H1, H2, H3 and H4X.

Yep, that's correct. I didn't consider the H1 and H2 as they are too old while the path to the H4x is needlessly complicated and more expensive than the alternative.

[synn]

From the get go I had little sense that you were convinced you needed a MF, and all this talk to test that notion at least gives you assurance you're on the right track.
I think it would be a great choice, and it is a different way of working that I think makes photography more pleasurable in some ways. I hope people can provide you with the samples for you to make your final pick.
I'm a product shooter myself and why I always look for the smaller sensor out of MF is the DOF :-) But I have been so happy with the results, I keep on with it.

When you decide, I would highly recommend Doug for your purchase, as you end up getting a part of Doug to take home. That's how dedicated and supportive he is. Which will prove priceless. Best of luck. I know a lot of people have learned one or two things from this thread, I know I have, and will be looking at a DP1/2/3 for my point and shoot travel camera real soon! :-)

Thank you. I do look forward to it too!

[eronald]

Why are Sigma landscape images so good? After inspection, I see no haze. Maybe the Sigma sensor is highly resistant to lens flare, and/or has low reflectivity resulting in less camera flare. And maybe the Sigma bodies are better baffled against camera flare. 

I would expect the Sigma sensor to be engineered for flare rejection (oblique ray rejection) , because crosstalk is highly destructive of color information. Conventional sensors do not have this issue.

Maybe someone here could shoot a test eg. a medium sized black square printed on white paper, with both a sigma and some other camera?

Edmund

SD1 metadata contains experimental data for flat field (luma and chroma) for 3 focal lengths at 4 aperture values each. The rest is to be covered by interpolation and extrapolation. The shortest length is 30mm, that is nearly 50 mm equivalent. It is not very useful for extrapolation for wider lenses, say, 24 mm equivalent. Also the data is obtained from select lenses and does not work very well with other lenses. Sigma are making a mistake by not providing end user with calibration procedures. Even their service facilities can't calibrate for a lens. The algorithms they use achieve nice results for those select lenses, but that can be improved grossly if red shift factor would be taken into account. In fact, the calibration quality and algorithmic improvement between pre-Merrill metadata and Merrill is very substantial; and the Merrill sensor structure is a gigantic step forward in terms of colour consistency compared to pre-Merrill. I think there is a possibility that current price drop is to clear the stock a little to make room for some new camera models. However I'm not so sure going to larger sensor format will not bring old issues back. Even with 1.6x crop Merrill the 3D-map of the sensor field looks ugly. If I would be working on the sensor first thing to consider is to improve noise so that 14-bit for at least 2 upper layers start to make sense. Second, a flat field calibration device for at least 2 colour temperatures, available at service facilities is a necessity. Third, lens mount rigidness and sensor-to-mount parallelism tolerance need to be improved (and that is from just a month of shooting).

[ErikKaffehr]

Hi,

What I think we may see is the effect of the slope of red transmission curve. If the slope is very steep, it will give good separation in red content. On the other hand it is possible that it will shift tones. As that curve is very steep small differences in the curve give large variation.

I also enclose some spectral plots of some spots of my skin (58 years old caucasian male with low UV exposure). You can see that the steep slope of the red on the Dalsa sensor seems to coincide with the steep slope on reflectivity of caucasian skin, this may also apply to other skin types.

Just to make clear, these sensivity curves are much different to human vision:


I sort of plan to dig a bit deeper into this when I get time…

Best regards
Erik

One thing I have noticed from using the IQ160 (compared to Nikon and Canon DSLRs) is that it seems to show a wider range of hues. It shows differences in hues that the dslrs do not. Also all colors seem to have more "depth" to them. I do not know if I am using the right terminology but what happens is that there might be three slightly different reds in a scene for example and the dslr shows them as the same red color (even when taking care not to clip any channel) while the IQ160 will show three, slightly different reds, just like in reality. I do not know why that is but I believe it happens with all colors.

[eronald]

Erik, I'm not feeling very smart so I'm having trouble figuring out your post.  could you rescale your two last curves (not the cones) so they match, maybe put one below the other? and could you provide labels so we know what they are.

Hi,

What I think we may see is the effect of the slope of red transmission curve. If the slope is very steep, it will give good separation in red content. On the other hand it is possible that it will shift tones. As that curve is very steep small differences in the curve give large variation.

I also enclose some spectral plots of some spots of my skin (58 years old caucasian male with low UV exposure). You can see that the steep slope of the red on the Dalsa sensor seems to coincide with the steep slope on reflectivity of caucasian skin, this may also apply to other skin types.

Just to make clear, these sensivity curves are much different to human vision:


I sort of plan to dig a bit deeper into this when I get time…

Best regards
Erik

[ErikKaffehr]

Hi,

Sorry! I am working on it!

Best regards
Erik

Erik, I'm not feeling very smart so I'm having trouble figuring out your post.  could you rescale your two last curves (not the cones) so they match, maybe put one below the other? and could you provide labels so we know what they are.

[ErikKaffehr]

Hi,

I had a hunch that shadow detail in my Sony Alpha 99SLT images are cleaner than in my Phase One P45+ images. Today I shot two quite comparable high contrast images, and try to push shadow detail. Exposure was quite similar on both, according to RawDigger.

Sony Alpha image has been upscaled to P45+ image size.
Actual pixel detail is enclosed below, P45+ left and Sony Alpha right.

Both images processed in LR5.3 RC

Best regards
Erik Kaffehr

[jerome_m]

From the two histograms, it seems that the Sony exposes 1/2 to 2/3 EV hotter, see the position of the peaks in relation to EV1-EV2.

As to the shadows sample: the Sony has less chroma noise, but lacks all detail. It would thus seem that the "raw" data from the Sony sensor already includes some noise reduction, possibly at the sensor level.