Friday, January 11, 2013

Nikon D800 Performance Test

Sensor gain is roughly 0.00324 * ISO (ADU/e-), based on 16bit linear TIFF converted by DCRAW.

ISO 200 gain, measured to be 0.6659 ADU/e-
ISO 1600 gain, measured to be 5.0424 ADU/e-


Method 1

ISO 100 readout noise: 3.02 e-
ISO 200 readout noise: 2.44 e-
ISO 400 readout noise: 1.94 e-
ISO 800 readout noise: 1.90 e-
ISO 1250 readout noise: 1.76 e-
ISO 1600 readout noise: 2.27 e-
ISO 2500 readout noise: 2.10 e-
ISO 3200 readout noise: 2.03 e-
ISO 6400 readout noise: 2.02 e-

Effective dark current at about 20 deg C: 0.077 e-/sec
Effective dark current at about 3 deg C: 0.0045 e-/sec

Method 2

ISO 100 readout noise: 3.97 e-
ISO 200 readout noise: 2.87 e-
ISO 400 readout noise: 2.51 e-
ISO 800 readout noise: 2.30 e-
ISO 1250 readout noise: 2.19 e-
ISO 1600 readout noise: 2.43 e-
ISO 2500 readout noise: 2.42 e-
ISO 3200 readout noise: 2.37 e-
ISO 6400 readout noise: 2.37 e-

Effective dark current at about 20 deg C: 0.356 e-/sec
Effective dark current at about 3 deg C: 0.0781 e-/sec



Note: The root of the two different method is that Nikon clips off the zero, and does not apply a positive bias to the image. This produces lots of zeros in the image, and only the positive side of the noise distribution can be sampled if the image is very dark.

In method 1, the histogram of the pixel values is fitted with a Gaussian. In all the readout noise tests, the Gaussian are all centered at zero. In the dark tests (20 min exposure time for 20 deg C, and 60 min exposure time for 3 deg C), the Gaussian clearly shifts to a positive center. The standard deviation of the fitted Gaussian is treated as the noise value.

In method 2, it assumes equal positive and negative histograms, and a standard deviation is computed from the real data plus an imaginary negative data. This tends to strongly overestimate the noise especially when the above Gaussian is shifted positively. On the other hand, this may be a more robust method when the noise is comparable to unity. In such a case, it is very difficult to fit the half-sided positive Gaussian.

For readout noise, I would tend to trust method 2.

Both method 1 and 2 were done in the image level. It considers the distribution of pixel values of the entire image all together. A proper way to measure dark noise is to consider variation at single pixel levels. To do so, a third method is employed to estimate noise.  In this method, series of simulated Gaussian distributions are clipped at zero and their mean and standard deviations are measured. The real measured mean and standard variation of each pixel are then compared with the simulated ones, to estimate the original (unclipped) mean and standard variation.

Method 3


Effective dark current at about 20 deg C: 0.067 e-/sec
Effective dark current at about 3 deg C: 0.0056 e-/sec


The results from method 3 and method 1 are more consistent. I believe they are more robust than method 2.  So the dark current of D800 should be around 0.07 e-/sec at 20 deg C ambient and 0.006 e-/sec at 3 deg C ambient.

Canon 5D2 performance summary

Based on this.

Readout noise at ISO 100: 28.92 e-

Readout noise at ISO 200: 14.59 e-
Readout noise at ISO 400: 7.71 e-
Readout noise at ISO 800: 4.19 e-
Readout noise at ISO 1600: 2.63 e-
Readout noise at ISO 3200: 2.55 e-

gain at ISO 200: 0.405 ADU/e-
gain at ISO 1600 3.2 ADU/e-

effective dark current at 20 deg C: 0.86 e-/sec
effective dark current at 3 deg C: 0.051 e-/sec