Comparing with Canon's older generation 6.4um pixel 5D Mark II, the 4.7um D800 shows notably higher SNR:
Saturday, March 24, 2012
DxOMark Prizes Nikon D800 Sensor
DxOMark, possibly the biggest sensor database in the world, proclaimed Nikon D800 sensor being the best one that DxOMark ever analysed. The 4.7um-pixel, 36.8MP full-frame sensor has shown 14.4 stops DR, color reproduction comparable with medium-format sensors and won score of 95 - the highest ever score in the database.
Comparing with Canon's older generation 6.4um pixel 5D Mark II, the 4.7um D800 shows notably higher SNR:
Comparing with Canon's older generation 6.4um pixel 5D Mark II, the 4.7um D800 shows notably higher SNR:
What does 14 stops mean?
ReplyDeleteOne stop is 2x exposure. For example, decreasing the lens' F# by one "stop" increases the exposure by 2x. So, the dynamic range of the camera - the ratio of the maximum signal to the lowest measurable signal, would be 2^14, or 20*log(2^14)=84dB... which is a very large DR for conventional pixel/sesensor
ReplyDelete-DP
this means that with 1e noise, the FWC should be 20ke. For a small pixel this is not bad.
ReplyDeleteThis is not a small pixel. A 4.7um pixel could have a FW of 50ke easily. With 2e of read noise, you get DR=88Db - they are in this ballpark.
ReplyDelete-DP
Please remember, DxOMark scores are normalized - the 14 stops applies to an image downsampled to 8 megapixels.
ReplyDeleteHow does down sampling affect the DR? Or you mean other performance matrices?
ReplyDelete* Anon wrote:
ReplyDelete> How does down sampling affect the DR?
By changing the analysis from a high spatial frequency to a low one. The raw data coming off a 36 MP DSLR is at a completely different spatial frequency than one from a 12 MP camera. Comparing these two numbers directly would only be realistic if one planned to crop out 24 MP and use just 12 MP from both cameras. But since that is not how most photographers use cameras in real life, one must normalize the spatial frequency difference between the two first in order to have a useful comparison. DxOMark chooses to do this by simulating how they would perform at the same 8x10 print size.
Comparing pixel performance without factoring in spatial frequency of the output display (e.g. print) is like comparing road noise of two vehicles both traveling at their top speed. Vehicle A has a top speed of 250 kph and very high road noise at that speed. Vehicle B has a top speed of only 100 kph and of course much less road noise. But no one would be foolish enough to claim that Vehicle B was quieter based on that alone. First one must measure the road noise of both vehicles traveling at the same speed (100 kph) to make a valid comparison. It may be that Vehicle A actually has less noise at 100 kph.
In the same way, sensor characterization must take at least some consideration for spatial frequency before conclusions can be drawn. Exactly what consideration is best varies by circustmances such as output display characteristics as well as viewer subjectivity. Downsampling is a very simple and reasonable method.
To measure the effect yourself, take a monochrome raw file and resample it with an anti-aliasing filter (or apply your own gaussian blur first to remove any detail that is above the new Nyquist of the resampled size). Then re-measure the RMS ADU of the same tonal level. You'll find that it is reduced compared to the full resolution file.
A simple way to think about it is that every raw file contains all spatial frequencies from zero to Nyquist and that noise power scales linearly with spatial frequency. Downsampling simply removes the higher spatial frequencies (with their higher noise power), leaving the lower spatial frequencies (and lower noise powers) in tact. Of course, there are far better ways of reducing noise power than simply chopping off resolution. But it is perfectly well suited for the purpose of normalizing data from different cameras.
There are a number of other factors that can come into play, such as something that changes the linear relationship between noise power and spatial frequency, such as noise reduction applied to raw data, or using a raw conversion that applies noise reduction. Another issue is the presence of non-random noise such as FPN, which is not helped as much by downsampling.
Hope that helps.