26.4mp 5D Mark III Mid-year? [CR1]

[davidpeter]

Dudes. The price of the AA filter and the chip itself grows exponential with both size and pixel density. Sooner or later, you have to choose. Haven't you recognised that the bigger sensors always had wider pixel pitch?

 

[gmrza]

If these specifications are true, I can see the D800 selling well.

Uh, perhaps you haven't been paying attention, but the factory where Nikon makes all of its pro-level DSLRs was seriously impacted by the earthquake earlier in the year. If that factory is responsible for the D800 then there could be a significant delay to it and other D? series cameras.

I wouldn't be surprised if this is the case and Canon knows it and is pouncing early because of it.

Canon has also been hit badly - I'm not sure how badly. The company I work for also owns a distribution business, which includes Canon. When I checked yesterday, there was one (1) lens in stock. There have not been any xD cameras in stock for a while as well.

 

[CJRodgers]

If the 5d mkiii had noise improvement would it make using f2.8 better for low light as you can use higher ISO?

 

[torger]

If the 5d mkiii had noise improvement would it make using f2.8 better for low light as you can use higher ISO?

Yes, but at some point there's a limit. There are two parts of noise, the camera's own ("read noise") and noise in light itself (photon shot noise). Most people forget about shot noise and assume that noise at high ISO is just contributed from the camera electronics. Actually, the shot noise is often dominating (I have not found out where the line goes, but perhaps some other reader of this forum knows).

The signal-to-noise ratio of the shot noise reduces the more light you gather, that is lower shutter speeds, larger apertures, larger sensor size. That is you always want to gather as much light as possible, to be able to use as low ISO as possible.

One other factor to know about is that not all f2.8 are the same concerning light transmission, and it is not only about vignetting. There can be quite large differences between lenses, some lenses can transmit a half stop less light than others at the same aperture. For example the 24-70 f/2.8 has a transmission corresponding to ~f/3.4 while a large aperture prime lens set at f/2.8 typically has much better transmission (did not find an example measurement unfortunately, but it can easily be observed in testing).

 

[NotABunny]

5DII high ISO performance is much better than the 7D

Tuggen didn't say that the 5D2's high ISO performance is the same as the 7D's. He said that the 5D2's high ISO performance is the same as the 7D's per UNIT AREA of sensor.

Of course the 5D2 outputs cleaner images than 7D since it has a 1.6 ^ 2 times bigger sensor, but each square millimeter has the same noise level

Ok, fine, but s/he is still wrong. There are two main factors that affect noise - sensor size and pixel size. Some will argue that pixel size alone determines noise, which is also incorrect. The total light-gathering capability (i.e. size) of the sensor is the primary factor, but smaller pixels do collect fewer photons per pixel, meaning less signal and a lower SNR. With a strong signal (i.e. good light), photon noise dominates and there is effectively no difference in noise from different size pixels. But as light levels drop and gain is applied, read noise has a greater contribution. In that scenario, the smaller pixels of the 7D are going to produce more noise per unit area than the larger pixels of the 5DII.

Okay, but this multitude of factors which affect the noise level is just theory. However, if you have physical evidence that the pixel size alters the noise level of photos, I really want to see it. I mean, physical evidence which meets scientific comparison criteria, not photos like those from IR ( http://www.imaging-resource.com ) where even the photos taken with a D3x and a D3s are not taken in the same light (the RAW images have different brightnesses: about 0.5 stops of difference).

Here is a simple way to do it: rent say a 1D4 and a D3s (or a D3x and a D3s, though they are separated by several years of technological advancements), take photos of the same subject, in the same light, with the same shutter speed, same F-number, and with the same output for the brightness of the photos. The only thing which may vary is the ISO (if the cameras do indeed have different ISO sensibilities); if any other parameter varies then that's like comparing apples and oranges. (Of course, this ignores the difference in the transmissivity of the lenses.)

Here is physical evidence that the pixel size does NOT alter (in practice, not in theory) the noise level of photos even at ISO 12800: http://www.canonrumors.com/forum/index.php/topic,255.msg3911.html#msg3911

Of course, one must really understand that sensor size normalization is necessary because we want to scientifically compare noise per UNIT AREA (not per photo since the sensors have different sizes). Since the 2 cameras have a pixel density per UNIT AREA whose ratio is 2.5, one can just see that pixel density doesn't affect noise per UNIT AREA. Consequently, pixel density is irrelevant in the current technological context.

 

[CJRodgers]

If the 5d mkiii had noise improvement would it make using f2.8 better for low light as you can use higher ISO?

Yes, but at some point there's a limit. There are two parts of noise, the camera's own ("read noise") and noise in light itself (photon shot noise). Most people forget about shot noise and assume that noise at high ISO is just contributed from the camera electronics. Actually, the shot noise is often dominating (I have not found out where the line goes, but perhaps some other reader of this forum knows).

The signal-to-noise ratio of the shot noise reduces the more light you gather, that is lower shutter speeds, larger apertures, larger sensor size. That is you always want to gather as much light as possible, to be able to use as low ISO as possible.

One other factor to know about is that not all f2.8 are the same concerning light transmission, and it is not only about vignetting. There can be quite large differences between lenses, some lenses can transmit a half stop less light than others at the same aperture. For example the 24-70 f/2.8 has a transmission corresponding to ~f/3.4 while a large aperture prime lens set at f/2.8 typically has much better transmission (did not find an example measurement unfortunately, but it can easily be observed in testing).

Yea i would love to know what the theortical best the 24-70L could be. This would be an ideal lens if it could handle low light just a bit better. So do you think no matter how good the noise handling in the camera, this lens could never be fast enough for really low light.

 

[torger]

Yea i would love to know what the theortical best the 24-70L could be. This would be an ideal lens if it could handle low light just a bit better. So do you think no matter how good the noise handling in the camera, this lens could never be fast enough for really low light.

It depends on what quality you want. For me personally, I often find f/2.8 a bit dark for indoor shooting. On the other hand, when going below f/2, the depth of field gets so short that it often becomes a problem (tricky to get focus right). Being able to go down to f/2 and sometimes below I think is often valuable indoors. If I know I'm going to shoot indoors I prefer a prime 35 or 50mm over a f/2.8 zoom - the wider aperture range gives me more flexibility than the zoom. The 50mm f/1.4 is a real good price/performance option, a bit narrow on APS-C though for being indoors.

 

[Loswr]

Here is physical evidence that the pixel size does NOT alter (in practice, not in theory) the noise level of photos even at ISO 12800: http://www.canonrumors.com/forum/index.php/topic,255.msg3911.html#msg3911

Do you mean the links to the photos on Juza's website? Those are 'fact'? In the long list of websites that purport to provide factual information, the last place I'd go is Ken Rockwell's. But the second to last place I'd go is Juza's. He does capture great images - but technical information? No. His reviews, at least, are filled with technical inaccuracies.

Here is a simple way to do it: rent say a 1D4 and a D3s (or a D3x and a D3s, though they are separated by several years of technological advancements), take photos of the same subject, in the same light, with the same shutter speed, same F-number, and with the same output for the brightness of the photos. The only thing which may vary is the ISO (if the cameras do indeed have different ISO sensibilities); if any other parameter varies then that's like comparing apples and oranges. (Of course, this ignores the difference in the transmissivity of the lenses.)

Your simple way sounds like exactly what Juza did in your link in the other post:

At http://www.juzaphoto.com/eng/articles/canon_1d_mark4_review_comparisons.htm there are 2 images (JPEGs converted from RAW) of a scene, taken in the SAME light, one with Canon 1D4, one with Nikon D3s, at ISO 12800....At the end, you'll have two images whose quality is indistinguishable.

That's a simple way to make a practical comparison - but if you're looking for 'physical evidence which meets scientific comparison criteria' then comparing a 1D4 and a D3s is a simple way to make a flawed and meaningless comparison.

First off, those are JPGs converted from RAW. Converted how? Different RAW converters handle different files differently. Even if the same RAW converter was used (e.g. ACR), different amounts of NR are applied to files from different cameras, and NR is applied even if the NR setting is turned off in the software. A real analysis would involve starting with the RAW files themselves (Juza doesn't make those available, just the JPGs), and analyzing the raw data itself using IRIS, Rawnalyze, or dcraw.

Even that's a flawed comparison - are you aware that RAW images are not really the raw data coming off the sensor following some sort of standard ADC? There is processing that occurs in-camera prior to the RAW file being written. On att least some Nikon models (D3, D300), a portion of the low-level signal is truncated during ADC, which obvoiusly occurs prior to the RAW file write. Astrophotographers using Nikons employ a 'mode 3 workaround' (modes 1 and 2 being Off and On) - mode 3 means powering off the camera during the dark frame exposure that follows the long exposure. Why? Because in addition to subtracting the dark frame prior to writing the RAW file, the camera also applies a median blur function to the image - something you might not want, but is baked into the RAW file creation in-camera. It's pretty likely that Canon also applies some processing of the image data in-camera prior to writing the RAW file. So comparing Canon vs. Nikon for noise in RAW files is really comparing apples and oranges.

Okay, but this multitude of factors which affect the noise level is just theory. However, if you have physical evidence that the pixel size alters the noise level of photos, I really want to see it. I mean, physical evidence which meets scientific comparison criteria...

If you'd like a reasonably cogent, scientifically-based discussion of image sensor noise, try the link below, in particular section 3c.

http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/index.html

 

[NotABunny]

First off, those are A real analysis would involve starting with the RAW files themselves (Juza doesn't make those available, just the JPGs), and analyzing the raw data itself using IRIS, Rawnalyze, or dcraw. Even that's a flawed comparison

Yes, those images are not ideal, but they're far better than anything else I've seen (and by the way, his interpretation of the photos is opposite - he's comparing noise at pixel level). It would be really great if you could show a better comparison.

But you're ignoring the practical point. Those images are taken at a very high ISO, where one sensor has 2.5 times the pixel density of the other. So, what does it take for all the factors that you list to cause a visible noise difference? ISO 1 million or a sensor with 100 MP? Does that make any practical difference today?

Even the RAW photos from IR taken with the D3x and the D3s, at ISO 12800, show very little noise differences (and only in the shadows, in favor of the newer one, of course), and one has twice the pixel density of the other. So, what does it take for the noise differences to be more than barely visible?

I'm not sure if we are even debating the same subject?! It makes no difference to me (or to anyone who screams at Canon to put only 10 MP on a FF sensor) if the noise differences can only be quantified with statistical means and aren't visible to the naked eye.


Later edit:
Actually, to make this simpler, I admit that a higher pixel density generates nosier images, but I want the people who want Canon to put only 10 MP on a FF sensor, to see what difference that would make: invisible to the human eye (unless you're looking for it).

 

[Loswr]

Does that make any practical difference today?

I think the real thing that obviates this discussion is that 'acceptable' (noise, resolution, etc.) is totally subjective. Some people are perfectly happy with ISO 800 on a digicam. Others find ISO 400 on a 5D a little too noisy. Some people will only print up to 8x10", where the downsampling from an 18 MP image will effectively reduce even comparatively high noise levels, whereas others (me, for example) have 24x36" prints on the walls, and want them sharp and detailed with very low noise. What looks sharp and clean in a 4x6" print can easily fall apart when the print covers a 36-times larger area.

 

[NotABunny]

Does that make any practical difference today?

I think the real thing that obviates this discussion is that 'acceptable' (noise, resolution, etc.) is totally subjective. Some people are perfectly happy with ISO 800 on a digicam.... whereas others (me, for example) have 24x36" prints on the walls, and want them sharp and detailed with very low noise.

You're right, "acceptable" is subjective.


Blabber:

I am not interested in resolution, either way. (Maybe if I had the possibility to use 30 MP, I would start cropping.) I am the kind of guy who shoots (processes and keeps) thousands of photos taken at ISO 3200 (from events). I am the guy who wants 3 stops less noise in his photos, but I have learned that the main cause of the technically bad look of a photo is not noise, it's light, not its amount, but its physical properties, like its spectral power distribution ((lack of) similarity with D65). Actually, I want to be able to shoot at ISO 12800, so that I can use my 70-200 F4 IS at 1 / 200 s.

If you are someone who thinks that noise is what makes your photo look bad from a technical point of view, you should experiment with your camera in broad daylight at ISO 3200 or higher to see what your camera is capable of. Then repeat with indoor lighting, like incandescent, fluorescent or sodium lights, or with dusk light. You should really ask yourself if its the fault of the sensor's noise or if it's the light itself that's just "wrong".

You can see here and here photos taken in normal room light at ISO 3200 (no flash, but a fluorescent tube, some sunlight and lots of luck for catching the right moment when the tube was in its best phase). Or here is one taken in daylight at the entrance of a hangar (this had lots of light, but still ISO 3200).

Does anyone believe that a lower resolution (how low? 2 MP?) would make the photos look better from a technical point of view? It would not. The problem with low light and high ISO is not the noise, is that the sensor can no longer give you dynamic and tonal range. It just can't do it because it has no light to "slice"; in other words, the tonal range quanta is an absolute amount of light, not 1/2^12 of whatever light is available. You can have whatever resolution and noise level you want, if you don't have lucky light, dynamic range, and most importantly tonal range, you just can't get a technically good photo.

 

[Loswr]

In the specific situation where you need longest possible reach higher pixel density is an advantage....There are no disadvantages with higher density.

Ahhhh...well, then...I hope you don't use a dSLR when you need reach, because it sounds like the Canon PowerShot SX30 IS is the perfect camera for that. 14 megapixels packed into in a 5.6x crop factor sensor for a very high pixel density, with a 150mm lens (840mm FF-equivalent), and no disadvantages. Well, gosh...it's just the perfect camera! I bet pro wildlife shooters are all selling off their 1DIVs and 600mm f/4L lenses, and buying SX30s, right? The SX30 even has IS, so the Gitzo and gimbal head can be left at home. Right?!?

 

[kirillica]

Yes, you are as wrong as in your previouslies posts.
In the specific situation where you need longest possible reach higher pixel density is an advantage. This has nothing to do with APS or FF. The problem is that FF bodies currently have lower pixel density than APS bodies forcing many to buy also an APS body (how often haven't I heard people adding a D300 to their D700 because of this reason only). As soon as FF pixel density reach same level as APS this will no longer be needed.
There are no disadvantages with higher density (the myth of lower high ISO nosie performace has since long been killed even though there are still a few with no knowledge that keep shouthing it) but there are advantages. Therefor this should be a requirement.

ahahaha.... simply read that neuroanatomist wrote and use powershot and mobile phones because they have no disadvantages. ;D ;D LOL, man, you made my day!

btw, good photographers use appropriate lenses and not a crop-factor to increase a focal length. if you need 800mm - use 800mm L lens with a 5d/1d or whatever you can afford, but promoting a powershot with 150mm in this case... kind a silly, isn't it?

 

[epsiloneri]

... it sounds like the Canon PowerShot SX30 IS is the perfect camera ...

With 1.3 µm pixels, a 150 mm lens would produce a pixel scale of 1.8 arcsec/pix, meaning a well-sampled image would have a resolution of 3.6 arcsec. This corresponds to an aperture of 28 mm, or an "f-ratio" of 150/28 = 5.4. The SX30 lens is 5.6 at 150mm, not too far.

Thus, if there was plenty of light, and the optics were diffraction limited (we wish, 35x zoom!), then indeed the SX30 could have been perfect.

The problem is of course that 28 mm is a very small aperture, collecting very little light. Scaling everything up by a factor of 4, we would have a 600mm lens (still ~5.6) with a ~110mm aperture and the same scale per 5.2 µm pixel (for a 14 Mpix sensor), but would now collect 4^2 = 16 times as much light, and be 4 times further from the diffraction limit.

We could also scale up everything with 4x except for the pixel pitch, resulting in a 224 Mpix sensor. Then the light gathering capability per pixel would be the same as for the SX30, but the pixel scale would be 0.5 arcsec/pix.

Conclusion: As long as you have sufficiently many photons and do not over-sample the resolution of the image, higher pixel densities result in higher resolution images.

 

[epsiloneri]

if you need 800mm - use 800mm L lens with a 5d/1d or whatever you can afford, but promoting a powershot with 150mm in this case... kind a silly, isn't it?

If the SX30 can do the job for 2% of the cost, why not? The problem is that it cannot, of course, because its optics is not diffraction limited at 150mm, and the light gathering power is abysmal. That said, if you increased the pixel density of 5D/1D then you would get higher resolution images ("longer reach"), so it does make sense to use a 7D in those cases. Which I think was Tuggen's point.

 

[epsiloneri]

Ahhhh...well, then...I hope you don't use a dSLR when you need reach, because it sounds like the Canon PowerShot SX30 IS is the perfect camera for that.

Why are you talking about PS when I'm talking about DSLR?

Perhaps I should let neuroanatomist reply that one to you, but since you don't seem to appreciates each others argument, maybe I can help:

1) You said there are only advantages to having higher pixel densities. That is obviously an exaggeration that you probably did not mean literally.

2) neuroanatomist gave the example of SX30 to show that a higher pixel density is not sufficient. He probably gave the example to make you think and realise why you are wrong in the "high pixel densities always rule" assumption. However, I don't think his example was very successful for that purpose because a) There are many other variables in addition to pixel density that are not kept constant (like sensor size, optics), so it's hard to from this one example to disentangle what the significance of the higher pixel density is. I tried to resolve it in a post above, but perhaps it was too technical to read well. b) He was not very clear with what the example was supposed to demonstrate. c) The tone in his reply was unnecessarily deprecatory, bound to fail an explanatory purpose and trigger the reaction you gave.

To be clear, here is an explicit list of some disadvantages with higher pixel densities (in approximate order of significance):

a) Readout noise increases with number of pixels
b) Slower readout time (limits your images per second rate)
c) More quickly gets limited by diffraction, so needs faster optics to be useful
d) Requires proportionally smaller tolerances for the camera house / optics to make use of the pixels
e) More sensitive to illumination direction (limb darkening)
f) Space/processing requirements increase
g) More expensive to manufacture

Note that I don't list noise as a disadvantage, since the quantum efficiency and collecting area does not change much with pixel density (only in the case where you are read-out noise limited will higher pixel densities produce a noisier image). Also, the dynamic range will not change significantly either, because the storage capacity of pixels is usually determined by their areas, so even if pixels are smaller and have smaller capacities, the number of photons they have to take care of is proportionally smaller.

The list of advantages with higher pixel density I can think of is much shorter, but very significant:

a) Potentially resolves finer detail in an image

I say potentially, because this statement is only true within certain limits. E.g., there is a limit how fine detail the optics will resolve. For small apertures, this limit may well be the diffraction limit (this is a physical limit); in general it is probably more common with imperfect optics. It also assumes that you have sufficient light and short enough exposure time for the pixels to be well exposed without introducing motion blur at the pixel scale.

There is always a balance between the advantages and disadvantages that sets the optimum pixel density. Depending on how much weight you give the different properties, this optimum balance will shift. For FF cameras with current optics, I think a practical upper limit on the number of pixels is around 50 Mpix, approximately the pixel density of 7D. Going beyond that does not make much sense to me, unless there is a revolution in lens manufacturing. I expect higher resolution images will be the domain of larger size sensors, medium or large format, as it is much easier to produce appropriate optics for them. I believe this will be even more true in the future, as sensors will be increasingly better and less expensive, while I don't expect manufacturing of optics to improve at the same pace.

I'll be happy to discuss any items above you may disagree with.

 

[Loswr]

However if they could make a FF with a, high quality, PS sensor density and image quality per area unit I would buy it since it would easily outperform 5Dmk2 , as I use now, at low ISO and have equal or better nosie and sharpness at high ISO.

Sorry, but I just don't buy that statement. For a given sensor size, smaller pixels will resolve more (potentially, provided they aren't limited by some other part of the system, e.g. lens resolution or diffraction), but will also be noisier.

Here's an example with both quantitative and qualitative information: Super Small, Sub 2μm Pixels for Novel CMOS Image Sensors, G.Agranov, R.Mauritzson; S.Barna, J.Jiang, A.Dokoutchaev, X.Fan; X.Li, in Proc, 2007 International Image Sensor Workshop, Ogunquit, Maine, USA June 7-10, 2007.

In the paper, they describe and test three sensors of the same size (1/4") but with different pixel densities and thus different sizes (from 5.6 μm, which is about the same as the 1DIV, down to 1.75 µm, which is a bit smaller than the S95/G12). They make the point that with the smaller pixels, they achieve increased resolution while maintaining similar sensitivity. If you look at Figure 7, the smaller pixels clearly offer a resolution advantage - no argument there. If you look at the lower panels (low light), it's also clear that with the same size sensor, as the pixels get smaller the image noise increases in light-limiting conditions.

 

[epsiloneri]

They make the point that with the smaller pixels, they achieve increased resolution while maintaining similar sensitivity. [...] If you look at the lower panels (low light), it's also clear that with the same size sensor, as the pixels get smaller the image noise increases in light-limiting conditions.

Thanks for the link to the paper. I glanced through it quickly. I find it interesting, though, that according to the authors, the image noise does not change with pixel scale. That seems to be their major conclusion, and is linked to maintaining sensitivity when shrinking pixel size. That seems at odds with what you say about the small-pixel pitch low-light sub-image in Fig 7 being noisier. I have two possible explanations:

1) I agree that the small pixel pitch image looks noisier, but it's hard to be quantitative "by eye" looking at what are probably strongly under-sampled images in print (with an unknown sub-sampling algorithm). It could be, that if you measure the noise at the same spatial scale, then all three examples would be equally noisy, i.e. show the same S/N (this is in fact what the authors claim). It would have helped if they provided a sub-image enlargement as they did for the 1000 lux case.

2) The authors focus on the sensitivity, or quantum efficiency (QE), and they show that the QE can be kept more or less constant with pixel scale (with their particular technology), at least down to 1.75 μm. They don't pay too much attention to read-out noise, however. For bright conditions, photon-shot noise and thus QE will dominate the S/N, while in faint conditions, the read-out noise will become important. It could very well be that the authors ignored that aspect, or deemed it insignificant, thus overstating their conclusion that "IQ and low light sensitivity of the sensors with these smaller pixels is comparable with a 5.6 μm pixel-based imager".

BTW, interesting that cross-talk seems to be such a big issue in CMOS sensors, I had no idea.

 

[Loswr]

They make the point that with the smaller pixels, they achieve increased resolution while maintaining similar sensitivity. [...] If you look at the lower panels (low light), it's also clear that with the same size sensor, as the pixels get smaller the image noise increases in light-limiting conditions.

For bright conditions, photon-shot noise and thus QE will dominate the S/N, while in faint conditions, the read-out noise will become important. It could very well be that the authors ignored that aspect, or deemed it insignificant, thus overstating their conclusion that "IQ and low light sensitivity of the sensors with these smaller pixels is comparable with a 5.6 μm pixel-based imager".

That was my interpretation.

 

[hallwal]

If you look at Figure 7, the smaller pixels clearly offer a resolution advantage - no argument there. If you look at the lower panels (low light), it's also clear that with the same size sensor, as the pixels get smaller the image noise increases in light-limiting conditions.

True, but towards the bottom of those low light images, the text on the large-pixel image is indistinguishable, but the text on the small-pixel image is quite readable. It seems that it would be fairly easy (noise reduction or some other filter) to make the bottom-right image look just like the bottom left image; it would be impossible, however, to do the other way around. In other words, I'd rather have the rightmost image (small pixel) in every example, including the low-light one.