Tuesday, January 25, 2011

D700 versus D300 – Testing Pixel Density


In my previous article, I found that light density (the actual amount of light falling directly on the sensor in any given interval of time) is your friend. Generally speaking, as ISO rises, so does noise. But if you stop down in low light and use longer shutter speeds, the noise rises much more than if you shot wide open. So that adds great value of larger aperture lenses for low light shooting.

So this article turns its attention to testing of pixel density. As I mentioned in the previous article, there has been a raucous debate on the Fuji Forum at DPReview.com this past week. An incredible amount of measurebating and some really unpleasant looking comparisons between a tiny sensor in a compact and a cropped sensor in a dSLR. The problem being that the crop is so severe that there are no pixels left with which to compete. The results look like crap to me.

So I thought I’d take the proverbial swing at it. I’ve been arguing that higher pixel density leads to more chroma noise … that’s the ugly blue and yellow blotchy patches that really ruin an image, far more so than simple grain. And this test will show me if that is true. Also, what about the detail advantages of the resolution making up for the bigger pixels of the smaller files? Well, to test that we will upsize the smaller file to match. A small bit of sharpening to compensate for the interpolation and we are back to perfect parallelism.


So I wondered how to run the perfect test. And it hit me … the D700 can shoot in crop mode at 5mp. This is the equivalent of a D300 with less than half the resolution. If I use a tripod and the same lens on both, I have the perfect test platform … two cameras with identical sensor generations and sizes, but with different resolutions, and hence with different pixel densities.

The only challenge remaining was to shoot equivalent images with every parameter perfectly matched. Even one parameter with a mismatch can call the results into question. And let me tell you that this is far harder than it sounds. I shot the test for two hours, three nights in a row. And tonight I finally got a pair of NEF files with which I can find no faults. Every parameter is matched. Note: With NEF it is not necessary to match the jpeg engines up, but I do that anyway just to be perfectly safe. The manufacturers do like to twiddle the files without telling us after all.

So, first we get two great captures. I take my little soapstone sculpture of a family of owls and I wrap it in a black boa. Plenty of detail in the statue and plenty of shadow detail in the boa. Let’s see how this goes …

Here are all the settings in one go with the D300 on the left an the D700 on the right. There is not a single difference except for sizes, and that difference is the whole point of this test :-)


Just so we know what is happening at the sensor level, the following diagram shows the life-sized ratio from the full frame sensor of the D700 to the APS-C sensor of the D300. The D700 shows a mask in the viewfinder when I shoot in crop mode, which looks a lot like this.


So … the ratio of pixel densities is 5mp to 12mp, which means that crops from the D700 are coming from images that have had pixels interpolated 2.4 times. This is above the recommendation for optical teleconverters by quite a stretch. It certainly has the potential to put the D700 at a significant disadvantage, just as the gang of four on DPReview would have us believe.

Web Sized Images

Before I show the crops, let’s look at the originals downsized with no processing other than the equalization of the conversion in ACR. Here is the D300 version followed by the D700 version.

Now, remember that these web images are very small. About a half megapixel, which means that there is no possible advantage to the D300 here. And I think the D700’s much larger pixels has improved shadow detail and definition here. It is simply a more pleasant image to look at tonally. One reason why I shoot the D700 almost exclusively these days.

Crops and Analysis

Crop 1
The bridge of the nose between the eyes is where I focused both cameras.

Crop 1 shows us that the D700, after being enlarged 2.4 times with a small amount of sharpening to compensate for the interpolation, is not only holding its own, but the cleaner pixels are in fact beating up on the D300 a bit. Right between the nose you can see that chroma noise and grain are taking their toll on the edges, yet the D700’s much cleaner pixels are handling it with aplomb.

Remember that you are looking at crops from 40” prints … but these are exactly the kind of cameras that are used to make 40” prints :-) … so the test is perfectly apropos. And what this shows me is that the cleaner pixels of the D700 will in fact beat 2.4 times as many smaller pixels from the D300. This is what has been the basis of my argument with the gang of four on the forum. I can even explain it, although it taxes my limited knowledge base to do so … but that gets saved for the end.

Crop 2
The bottom right corner is the darkest overall part of the image.

Here, the difference is blindingly obvious. More grain, larger grain and out of control chroma noise on the D300. Much less of everything from the D700. So what happened? Smaller pixels … that’s what.

Crop 3
An area on the left side where a few hairs from the boa are standing up to be counted.

Again with the noise difference. Getting old by now and is expected. But again we find the D700 outperforming the D300 in details as well. This is counter intuitive until you realize that the issues here are mainly down to edge destruction by grain and chroma noise. There is zero noise reduction in these, so that cannot be involved. It’s down to the effects of noise causes by smaller pixels.

Crop 4
The left eye area.

Same thing here … larger grain, more grain, lots of color blotching. The result is that the 5mp are winning the detail battle again. The D700 makes a far better enlargement. This is reminiscent of the days of the D2Hs, a professional quality 4mp camera that was laughed at. Until, that is, people saw enlargements made from its stunningly clean pixels. This was the precursor of the D3 success and up to about 800 ISO it was pretty magnificent.

Well, it is obvious to me that shooting my D700 with the 18-200VR at 5mp is going to give me better results in high ISO situations than the D300 at full resolution. Who’d a thunk?

Final Analysis and Explanation

So How the heck can the D700 win a battle where it comes in with 2.4 times fewer pixels? Well, I am going to go out on a limb and suggest the following:

The D700’s pixels catch 2.4 times the number of photons, assuming rectangular pixels. The high ISO I am shooting amplifies the photons and the shot noise by 32 times in the analog domain. That’s the first  major imbalance tilting towards excess noise in the smaller pixels.

It also adds significant amp noise, which is the second imbalance towards noise. My theory is that the extra noise from both factors (assume read noise is constant on the same sensor generation) overwhelms the much smaller pixel counts, leading to much increased grain. This is clearly visible.

But how to we further explain the significant increase in chroma noise? The color blotches are very obvious after all. Well, the much smaller pixels in shadow are are going to starve a bit in the D300. And since this light is predominantly red, it is the red channel that is underrepresented. When we make the blue channel push to set white balance, again the small pixels are overwhelmed by the push of the blue channel, which is very significant. The green channel also takes a beating, which shows up as yellow blotching. This is the third major factor tilting to noise.

It works for me at least …

So … I’m not saying that we should all beg for fewer pixels. But if we can see the differences so clearly at these low pixel densities – D700 is 1.4MP/cm2 and D300 is 3.3MP/cm2 and with a 240% advantage for the D300 in resolution, then what exactly should we expect when we take pixel densities from the 45MP/cm2 in the ZS3 to the 50MP/cm2 in the ZS7? And with only a small advantage in resolution of 20%?

A bit of a travesty, I imagine … which is exactly what many have observed in the compacts, and which has now been confirmed in full by this test.

No comments: