D800 Noise and Pixel Resolution Question.

I could be wrong but I don't think anyone knows for sure at this time.

While 36mp seems massive, it is basically the same pixel density as the D7000's resolution, just a larger sensor size (DX vs FX). The d7000 performs well at ISO 6400 so there is no reason to think the D800 won't perform equally as well, if not better being that the D7000 is 1 year old technology.

Yes. A resizing operation normally is implemented by first interpolating the image, then low-pass filtering it (this is similar to averaging) and then decimation which is picking the required number of pixels from the low-pass filtered image at equal intervals on both axis.

Stars will normally occupy multiple pixels in the image so they aren't going to disappear (they might be slightly lower in contrast) when and reducing the 36MP image to 12MP. Using a 36MP image with size reduction in post should result in a slightly better quality image than using a 12MP image in the first place (if we assume noise per area is comparable) because a 12MP camera image is an interpolated 12MP whereas the 36MP capture resized to 12MP is interpolated at the 36MP level and decimating to 12MP leads to an image where each pixel has full color information and is likely to render pixel-level details somewhat better. The other thing is that the antialiasing filter in a 12MP camera attenuates the fine detail to some extent whereas the AA filter of a 36MP camera is thinner so it operates at higher spatial frequencies in the image. A D800E would be best for star fields as it has no AA filter at all. That should result in extremely crisp shots of stars.

The D800 has a single sensor with a Debayer, in essence the camera calculates groups of pixels, an example, 1 green pixel, 1 red pixel and 1 blue pixel would create a "super" color pixel. The Expeed processors then use mathematical algorithms to triple the actual pixel count. Because a 36MP color image has to have a 36MP Red image, a 36MP Blue, and a 36MP Green image. So a single 36MP chip cannot get a full color image without a computer calculating what 2/3rds of the image should look like. Based on this, I would theorize that 12MP would be slightly sharper, slightly more accurate color rendition; overall slightly higher IQ than 36MPs, because the computer isn't creating any pixels, its simply gathering 12MPs of Red, 12MPs of Green, & 12MPs of Blue, totals 36MPs, which means that the image is being creating directly from pixels on the sensor, and not from the computer calculating 2/3rds of the image. However I'm only theorizing. In theory only using physical pixels that are actually on the sensor recording the image from the lens will render more accurate results than a computer "creating" 2 additional pixels for every 1 pixel on the sensor.

Regardless of how pixels are debayered, rearranged, duplicated etc, they are still the same physical size and have the same physical light gathering abilities. By using different arrangements of pixels (12MPs vs 36MP) may have an effect on noise, but it will not have anywhere near the effect as getting a camera with physically larger pixels (D700, D3, D3s, D4).

Historically, I believe it's the case that cameras with larger pixels have handled noise at the image level better than cameras with higher resolution sensors - that is, a bigger photosite is better than binning smaller photosites. That's probably no surprise given that the overhead of each pixel cuts slightly into the sensor area, so (for example) I believe that two pixels from a D3x don't have the same sensor area as one pixel of the D3. Note that I'm talking about the sensor site here, not the area covered by microlenses, which by rights should be similar (every generation of camera for several years seems to have claimed that they've eliminated dead area between the microlenses of the previous one; I get very suspicious whenever they're called gapless).

That's not authoritative, but by way of example, I believe the consensus is that the D3/D700 are considered a superior low-light camera to the D3x or 5D2. The per-pixel low-light handling is obviously better, but I'm led to believe that the apparent image quality is better too.

All in all, I'm not expecting the D800 to be better than the D700 in low light, because the D7000 isn't - although it's very close. I'm expecting it to smoke the D700 in good light, however.

As for star fields, I would expect the majority of light from a (dim) star to hit one sensor site, if the lens can keep up, on a D800E - I suspect the D800E would kiss goodbye to knowing what colour the star was, however. I would expect starlight to be relatively reliable at hitting a very small number of sensor sites on any camera with a low pass filter. Downsample the image and the stars will, indeed, get dimmer (just as they would if you deliberately missed focus) - some of the sites contributing to an output pixel get no light, and the average is dim. Of course, if the light on the small number of pixels is very bright (and within the range of the sensor), it may still make a significant image contribution. If starlight didn't act as a point source, the observation that a 500mm f/4 shows more stars than a 14mm f/2.8 wouldn't be true, and we could all buy cheaper telescopes.

I'd be astonished if the D800's downsampling didn't bin and improve the noise behaviour, although it might do some decent filtering to retain a bit of sharpness, if we're lucky - the exact details won't, I'm sure, be known until the camera ships (but you can always do this in an image editor). I'm vaguely curious as to whether it might have a low resolution RAW format like some Canons - I might know when I don't need 36MP, but still think I've screwed up the exposure/white balance and want some more bits of accuracy. Especially if it might get above 4fps as a result.

Mike, why would you choose to reduce the pixel count in camera? You'd just be using the comparatively limited processing power of the camera's CPU to throw information away; when it's surely speedier to simply keep all the image data and process it to a smaller image in post using a more powerful computer, if you think it necessary.

Skyler, I'm not following your logic of "pixel tripling" above. The standard Bayer pattern actually uses a group of 4 photosites - RGGB - as the source of chroma information for each and every pixel. The "averaged" colour of the RGGB cluster surrounding the target photosite is calculated and applied to that pixel. Then the 2x2 matrix is stepped on one pixel either horizontally or vertically and the process is repeated. In this way each and every photosite becomes an RGB pixel; there's no real interpolation or "computer calculating 2/3rds of the image" happening, since each photosite/pixel is allocated a unique chroma value based on the combination of it's surrounding RGGB cluster.
All four of those Bayer-filtered RGGB photosites can have a totally different RGB value because a different 2x2 RGGB matrix is used to calculate each one. Notice the word "calculate", rather than "create" or "guess at". True this results in a lower chrominance resolution than luminance resolution, but it's still not interpolated in the sense of being created out of thin air.

Likewise, any higher noise due to smaller individual photosites being used will be averaged out and should result in a similar noise level to an equivalent physical area of larger photosites. It's only when pixel peeping at 100% that any higher noise will become apparent. In like-sized prints, or at a like-sized screen magnification the noise will appear very similar. However, the foregoing doesn't take into account the almost inevitable improvement in noise performance that comes with every new generation of sensor.

there's no real interpolation

What you are describing in that paragraph is precisely interpolation.

it's still not interpolated in the sense of being created out of thin air.

Inter means "between". You calculate the in-between values based on the known surrounding values.

Rodeo, of course you're right. It started as a thought as to how big/quick the buffer was and I wondered if the buffer fills with RAWs even when you're taking JPEG and where in the processing chain the JPEG is produced? Can you get more small JPEGs before the buffer fills? It was more the option to, rather than the wish to!

RJ, Skyler, I'm sure you both know what you mean, but for the sake of anyone confused about how debayering works, I wanted to try to explain the disparity in what you're saying.

With a (conventional) Bayer sensor, half the sensor sites record green information, one quarter record blue and one quarter record red. As such, it's true to say that the red or blue resolution of the image is one quarter of the total sensor site count and the green resolution is half the total sensor site count - and if you looked at a purely black-and-primary-colour image (if the colour matched the filter), that would be a pretty accurate assessment. "Debayering" interpolates the colours between sensor sites to create a full colour image with higher resolution; the most trivial way to do this would be to average (bilinearly) each channel between the samples of the associated colour. If you do that, you get, as discussed, an effectively lower-resolution image. Sigma are fond of claiming that this is the representative resolution of a Bayer sensor when discussing the merits of the Foveon sensor.

However, this simple image reconstruction approach assumes that there's no correlation between channels. Typically, that's not true - most objects aren't a pure colour, and, for example, knowing that you're following a greenish-yellow edge means that you can derive useful luminance information from primary-coloured sensor sites. This only works if there's enough information to correlate the channels, which is why a sharp star hitting a D800E is going to be difficult to colour correctly, and why moiré false colours are a problem.

So:

• Each sensor site only records one colour.
• Each output RGB pixel can contain a different value for each channel, since the offset from the samples is different for each pixel.
• Simple interpolation would give you the appearance of an up-scaled lower resolution image.
• The correlation between channels in real images means that, in practice, the recorded value at each sample can be used to derive luminance changes which affect all the channels at each sample point.
• The camera (or computer) is still "making up" information at each sensor site, but most of the time it's got quite a lot of relevant information with which to make an educated guess - if this wasn't true, the Bayer sensor wouldn't be so universal. Wikipedia's "demosaicing" page is a good starting point on this.

All of this explains why cameras tend to output YUV420 by default (you can make some deductions about the luminance of each pixel, but trying to deduce the colour of each sample is probably unreliable enough to make it not wasting JPEG image size.) Incidentally, I'll be curious as to whether it's possible to save small JPEGs with YUV4:4:4...

I hope that helps someone. :-)