Thanks so much to my buddies, Bob, Scott and John. We’ve all been buying and looking at various mirrorless systems and my buddies have neatly split into three camps. NEX (Scott), Fuji (Bob) and Samsung (John). So what is a geek to think?
Well the main issue is how good the products really are. They are all quite similar from the external view. All are 16MP-24MP with APS-C sized sensor and are interchangeable lense. The big differences are the sensors themselves. And we’ve been debating the very confusing measures of performance based on how well the sensor perform in low light (low noise is critical here), how well they do dynamic range (good for landscape where you have big difference between light and dark) and of course how much they resolve detail. The confusion is that there are lots of different test methodologies and camera vendors are actually doing a lot of signal processing in the background, so you have to ask if the image they produce digitally processed or not. So here is a summary of the debate…
DxOMark’s measurements seem to track other accurate measurements of dynamic range and noise. It’s a very good metric. For instance, Bill Claff has done some really good work. His methodology is extremely solid and was developed independently before DXO. He works directly with the Raw data. In fact, his noise measurements for Nikon are based on the optical black pixels that are used for calibration in most arrays, eliminating the Raw process.
To understand noise in a digital camera, you have to account for several sources:
Photon noise – it is as random as random gets and is
Read amplifier noise – Amplifier noise is also pretty darn close to random Gaussian.
Analog to Digital Convertor noise – generally random, unless the ADC has design issues
Bad design noise – an example of this is the column readout noise seen in most Canon sensors, which shows up as streaks across the image
ADC offset – fixed offset variations in the readout across sensor columns, seen as darker or lighter areas in the image
Most modern designs eliminate the non-random noise in the amplification change. Non-linear jumps, are usually due to programmable analog and digital gain. For example, on the D800 for best dynamic range and noise performance in low light environments, you want to shoot at ISO 1600, which is the last ISO level that does not use digital gain, to allow the full use of the LSB in the ADC. Amplification can then be done in post-processing with your own software rather than what is done automatically. Every modern camera has a certain “native ISO” which is the last level before there is digital gain.
[As as aside, this is a pretty complex topic that stackexchange covers but in short, for older Canon cameras (7D, 5D2, 600D and older), ISO 1600 is the last native level so you don't need to shoort more than that. Canon then uses analog amplifiers to get to ISO 3200 and then digital amplification to get to above ISO 6400. Put another way, with these cameras, you should stop at ISO 1600 and process the rest digitally on your computer. Later Canon's have better analog amplifiers so you can get to ISO 25600 so you can shoot better. As an aside Sony Exmor sensors (used in Nikon and of course Sony) use a different sensor design. They work up to ISO 12800 natively but don't do as well in extreme ISOs vs Canon. On the other hand, they have barely any noise at ISO 100-400 so dynamic range is amazing. The main thing though is that it is hard to figure out what that transition to digital gain is. Scott says it is ISO 1600 on his D800, but this last piece implies it is a little further from that. Perhaps ISO 12800. Confused. Me too!]
If the sensor and image processing engine are well-designed, then offset errors and column readout noise are extremely low, leaving random photon and amplifier noise. Thus the reason why downsampling and averaging are valid techniques. Looking at cameras like the D800 and D4 and comparing to theoretical best case photon noise performance shows them to be about 1.2 stops away from the theoretical best possible noise performance.
Fuji XTrans claims 1 stop better noise performance than the D7000. Given that the D7000 is 1.25 stops away from theoretical best dynamic range with only photon noise, that suggests that there is some “slight of hand” in Fuji’s claim. Removal of the anti-alising filter provides some of the improvement, since the filter itself creates loss within an octave of Nyquist (as can be seen when comparing the D800 to the D800e.
As for eliminating Moire, that’s crap. What they do is to apply a DSP filter across a much larger array section than a “normal” Bayer demosaicing algorithm. The filter will generally take care of color aliasing, due to the large number 2D samples being averaged. There will be errors that can be seen with pixel peeping, but across the image they will be randomized, eliminating the appearance of false color. What still remains is luminance Moire of repeating patterns at or above Nyquist. That’s just a fact of the physics. True Moire-free sensors will need a much higher pixel density that takes advantage of lens diffraction as the anti-alias filter.
Finally, I ran into this comparisons of out of camera jpgs from DP review. The interesting thing is to look at the resolution in the circular proportional scale on the right of the photo. Clearly there is something going on with the Fuji image processing. Looks like a bunch of averaging going on, which effectively makes for a low pass filter.
Having looked at a few photos taken with the X-Pro 1 I’d say there is a nice “creamy” quality to photos taken. They seem to sacrifice resolution for that smooth look, which isn’t bad. I don’t know about the lens system, but if you’re looking for a camera that puts out good looking jpgs without post processing, that’s on the par with a Nex5, this isn’t a bad choice. If you process RAW files, then I’d probably go with something else.