So what's the deal with pixels. As could probably have been anticipated, something of a quasi-religious war has broken out in some internet forums as supporters of the Nikon D800 (36.3MP) battle supporters of the Canon ESO 5D MkII (22.3MP) with cries of "Pixels Count!", "No they don't", "Noise matters" and "No it doesn't" ringing out across the land (well, across the forums anyway).
Where does the truth lie and is there any truth that's tired of standing?
Let's look at three sensors. Let's say two are 36mm x 24mm and one of them has 22.3MP and the other has 36.3MP. The other one is 23.2 x 14.9mm (Canon APS-C) and has 18MP. What does this imply?
|Canon EOS 5D MkIII||Canon EOS 7D||Nikon D800|
|Number of Pixels||22,300,000
(5760 x 3840)
(5184 x 3456)
(7360 x 4912)
|Sensor Size||36 x 24mm||22.3 X 14.9mm||36 x 24mm|
|Pixel pitch||6.25 microns||4.3 microns||4.89 microns|
|Maximum possible pixel size||6.25 x 6.25mm||4.3 x 4.3mm||4.89x4.89mm|
|Maximum effective pixel area||39 sq. microns||18.5 sq.mm||23.9 sq.mm|
|Maximum theoretical resolution||80 lp/mm||116 lp/mm||102 lp/mm|
|Max Print size at
|28.8 x 19.2||25.9 x 17.28||36.8 x 24.56|
|Max Print size at
|19.2 x 12.8||17.28 x 11.52||24.56 x 16.37|
It's instructive to note that despite the rather large sounding difference between 22.3MP and 36.3MP, the difference in resolution is only around 20%, The 22.3MP full frame sensor has 80% of the resolution of the 36.3MP full frame sensor. The 18MP APS-C sensor has the highest resolution of them all, about 40% more than the 22.3MP full frame and 15% more than the 36.3MP full frame. Of course it doesn't cover as much area as the full frame sensors, but it does "test" lenses to a greater extent due to its smaller pixel pitch and higher resolving ability. This means that any lens that looked good on an EOS 7D will look good on a Nikon D800 (if it could be mounted), at least over the APS-C equivalent area of the frame. This counters the argument that you'd need really, really good lenses to take advantage of a full frame 36MP sensor. It is true though that some lenses (e.g. mid priced wideangle zooms) might not be able to deliver enough quality at the extreme edges and corners to fully exploit the resolving power of the D800 sensor (though that may well be true to a lesser extent for the EOS 5D MkIII sensor too).
It should be noted that the resolution in the image not only depends on the native resolution of the sensor as discussed above, but also on the strength of the anti-aliasing filter used in front of the sensor as well as the quality of the lens used on the camera.
Now all else being equal (which it rarely is, but which makes comparisons easier), the larger the effective pixel area, the lower the noise and the lower the noise the higher the dynamic range. There's no simple relationship between pixel area and noise because there are multiple noise sources including shot noise (photon noise), readout noise and pattern noise. However you can sort of guestimate that if you double the pixel pitch (increase the pixel area by a factor of 4) then the signal/noise ratio improves by a factor of two. All else being equal. Which usually it isn't of course...
Dynamic range follows noise because dynamic range measures the difference between the weakest and strongest detectable signal. More noise makes it harder to detect the weak signals, so the dynamic range is smaller.
Lower noise also means that you can shoot at higher ISO settings and still get low noise images. You will see from the examples above that the manufacturers specified ISO range is highest for the sensor with the largest pixels, the EOS 5D MkIII. 2 stops higher in fact for both the native ISO range and for the expanded ISO range.
The maximum possible resolution number sin the table are based on the Nyquist sampling theory which says that if you sample a signal at twice per cycle of the highest (sinusoida) frequency component present, then you can fully and unambiguously reconstruct the original signal. That doesn't actually mean that you can resolve detail that's as narrow as two pixels since the signal falling on the pixels isn't likely to be a sinusoidal waveform, or indeed a waveform in which the highest frequency component is two pixels wide. However it gives us a rough estimate of the relative resolving power of a sensor based on the pixel pitch, i.e. the distance between pixel centers. What is true is that the smaller the pixel pitch, the finer the detail that can be recorded - assuming that detail is present in the image in the first place.
So what you get with sensors is a tradeoff of resolution for noise (and the related dynamic range). Big pixels give you lower resolution, lower noise and higher dynamic range. Small pixels give you higher resolution, higher noise and lower dynamic range. The real question is what constitutes "High Enough" resolution and "Low Enough" noise. Can you have both?
For large prints, most professional printers would agree that you need about 200 pixels per inch to produce excellent quality images. For smaller prints which are viewed from a closer distance, the best quality might require up to 300 pixels per inch.
From the table above you can see that all three sensors should be capable of making excellent quality prints which are larger than most people will ever print. They are all good for exhibition quality printing at a size of 20" x 24" (a standard size). This is with standard files. There is software out there that is specifically designed for making large prints and it can just about double the size of the image without an appreciable quality loss (again for large prints which will be viewed from several feet away). I've seen stunning 4ft x 6ft prints made from ~20MP cameras. Examples of software used or upsizing are are "Genuine Fractals" (which has changed its name to "Perfect Resize") and Alien Skin's "Blow Up". So realistically the number of photographers who will be pixel challenged by any of the three examples above is small. The one exception to this might be if the photographer wants to make a large print from a highly cropped image. In that case the 18MP APS-C sensor might be the best choice from a resolution viewpoint because it has the highest native sensor resolution and, if using the same lens, would require less cropping than either full frame sensor.
Even without the help of software, all three cameras are capable of super high quality exhibition 11x14 prints (even with some cropping) suitable for close viewing, and that's about as large as most people ever print. In fact they'd all probably still make excellent 16x20 prints.
Video doesn't need a high resolution sensor - and in fact a high resolution sensor can actually be a handicap. 1080HD video is only 1920 x 1080 pixels. That's all you need. You gain nothing by downsizing a higher resolution image (as all these cameras have to do), in fact you lose because of several factors. The first problem is downloading all the data from a high resolution sensor at the required video frame rate (up to 30 frames/second). Some cameras do this by skipping lines of pixels (rows or columns depending on how the sensor is designed). This means less data needs to be transferred, but it can result in worse diagonal "jaggies" ("staircasing") due to the missing data. The second problem is Moire fringing from downsizing. Technically this results from undersampling the larger image which results in aliasing issues. It you don't skip lines the aliasing and Moire patterning is reduced, but the ideal would actually be a 1920 x 1080 pixel sensor. No aliasing and faster readout. Look at a dedicated video camera like the for example. It has a 1920x1080 pixel sensor measuring 23.6 x 13.3 mm. No downsampling required.
The Canon EOS 5D MkIII does not skip lines. It reads out all the pixels before downsampling. This results in less Moire patterning and aliasing effects than were present in video from the EOS 5D MkII (which used a line skipping technique). My understanding right now is that the Nikon D800 uses line skipping, though I don't have direct confirmation of that through any Nikon source.
A larger sensor does allow more flexibility in video format, and even an APS-C 18MP sensor could easily generate a 4K video signal which is at most 4096×3112 pixels (some variations on the 4K format have lower pixels counts). 8K video is 8192×4320 pixels and is used for some professional movie mastering applications. At the moment even 4K video a moot point because (a) no DSLR generates video with higher resolution than 1080HD and (b) there are no production consumer devices (TVs) capable of showing 4K or 8K video anyway (though 4K prototypes have been shown).
Just to show that Canon probably could make a high pixel count full frame sensor if they wanted to, a couple of years ago they reported on the development of a 120MP full frame sensor which can deliver HD video from any part of the sensor. Only 1/60th of the sensor area is needed for a full 1080HD video image! Canon have announced no current intention of using this in any commercial product, but clearly they don't build these things just because their engineers want to have fun. In fact they have shown an engineering prototype of a camera using this sensor (but with no current plans to put it in production) and have talked about a far future camera based on a sensor like this which only shoots video (but 120MP per frame video...) from which stills and digital zooming could be produced by just looking at a subset of the pixels in a video frame. 2020? 2030? who knows if and when.
You might think that a sensor like this would have such small pixels that the noise would be very high, but in fact the pixel area would be about twice that found in today's 14MP digicams so the noise should be lower than those cameras show. Not great but not terrible, especially if you stay below ISO 800.
The bottom line is that neither the D800 nor the EOS 5D MkIII sensor is "better". They are different. The Nikon sensor should have higher resolution which will be desirable for those making very large prints or who need to significantly crop their images. On the other hand the EOS 5D MkIII sensor will produce images with lower noise and higher dynamic range in lower light conditions where the use of higher ISO settings are desirable. For smaller prints at lower ISO settings (which is where most amateur photographers will be working most of the time), the sensor pixel count and noise characteristics won't matter.
It should be remembered that the Nikon D800 has larger pixels then the Canon EOS 7D and the 7D isn't exactly a noisy camera! So despite the high pixel count the noise level of the D800 will probably rate as very good. maybe even better than that of the EOS 7D. However the EOS 5D MkIII is said to be a couple of stops better than the EOS 5D MkII (at least for camera JPEGs), and the 5D MkII is maybe a stop less noisy than the EOS 7D, so it looks like the 5DMkIII will probably have a couple of stops lower noise then the D800. That will be of no importance at all at ISO 100, but of significant importance at ISO 6400.
Some photographers will prefer the Nikon approach, some will favor the Canon approach. It's nice to have choices! and of course there's FAR more to a camera than noise and resolution. Aspects such as AF, metering, ergonomics and features (aka "bells and whistles") are equally important - as is cost.
What would I buy if I didn't have an investment in Canon lenses? I'm not sure, but the $500 lower price and built in flash might make me lean slightly towards the Nikon D800, but it's only a slight lean!
Neither the Nikon D800 nor the Canon EOS 5D MkIII are shipping yet (03/04/12) but Adorama is taking pre-orders and cameras will ship on a "fist come, first served" basis.