Noise and DR measurements

[Guillermo Luijk]

Thanks to some RAW files contributors, I did some noise and dynamic range measurements of the Canon 5D, 5D Mark II and 7D at ISO100 and ISO1600.

Using Gabor's Rawnalyze I obtained the mean (Signal) and StdDev (Noise) for several patches. After correcting by the Canon's black offset and converting to EV RAW exposure values referred to sensor saturation, taking into account the particular saturation point for every camera at each ISO, and assuming sensor linearity, I plotted the following SNR curves:



The 5D curves match quite well those obtained in the past by Email Martinec and Gabor Schorr.


However those are per-pixel SNR curves. To make a fair comparision between the three cameras, and calculate comparable dynamic range values, I normalised all the SNR data to match the Canon 5D resolution. Done that the new SNR curves are as follows:



Quick conclusions:
 * The 5D2 beats the other two cameras SNR in aprox. 2/3EV for ISO100 and ISO1600, and presumably for any other real ISO
 * The 5D and 7D perform quite similar in terms of noise, what makes think that the improvement in technology compensated for the format size
 * At ISO100 the 7D has a slightly better SNR than the 5D in the deep shadows (dynamic range), but in well exposed areas the 5D is slightly better
 * At ISO1600 the bigger size of the 5D sensor provides a better noise performance than the 7D at any exposure level


By taking the SNR>12dB criteria for Dynamic Range, the following figures were obtained from the previous plot:



Quick conclusions:
 * The DR of the 5D Mark II is better in around 1/2-1EV at ISO100 and 1EV at ISO1600 than those of the 5D and 7D
 * The 5D and 7D have similar DR, slightly superior in the 7D at ISO100 and in the 5D at ISO1600
 * The reduction in DR capacity when raising ISO, is more notorious in the smaller sensor size


About the noise and SNR measurements, I was a bit surprised that for higher exposure values the measurements started to have quite a lot of variance. I wonder why was this since SNR should improve, and therefore be more stable between the different measured values, than with low exposure values. Or I am missing something?

The original article (SP witn online translation) here.

Regards

[Ray]

Interesting results, Guillermo, and thanks for showing them. It's surprising that the 7D is marginally better than the 5D at base ISO, after downsampling to the same resolution.

The 5D2 is clearly the best, as expected, but having resisted buying one so far, I think maybe I should wait for the next upgrade. I guess my Nikon D700 now sets the standard for low noise, from my perspective, and I wouldn't like to feel I was going backwards in any respect   .

[mike.online]

after picking up my 5dmkii yesterday, i'm happy with these results

i really wondered about this, and wanted more opinion than the DxO results, thanks

[Bart_van_der_Wolf]

About the noise and SNR measurements, I was a bit surprised that for higher exposure values the measurements started to have quite a lot of variance. I wonder why was this since SNR should improve, and therefore be more stable between the different measured values, than with low exposure values. Or I am missing something?

Hi Guillermo,

I haven't seen the underlying data, but I must assume that photon shot noise is the largest contributor to your basic data at increasing exposure levels. If the noise floor is relatively stable, and the higher exposure levels show a larger variance, then SNR is affected.

Cheers,
Bart

[Guillermo Luijk]

Interesting results, Guillermo, and thanks for showing them. It's surprising that the 7D is marginally better than the 5D at base ISO, after downsampling to the same resolution.

The 5D2 is clearly the best, as expected, but having resisted buying one so far, I think maybe I should wait for the next upgrade. I guess my Nikon D700 now sets the standard for low noise, from my perspective, and I wouldn't like to feel I was going backwards in any respect   .

The D700 is supposed to be a bit better than the 5D2. If you want (you or anyone having a D700) I can incorporate noise measurements from D700.

Just 2 or 3 RAW files at about 3EV intervals shot over a colour chart are needed, including well underexposed shots, and one extra blown RAW file. The only important thing is to have all patches very uniformly lighted (set light sources far from the chart, avoid lens vignetting, and blur the lens a bit not to capture the card's texture). Both for ISO200 and ISO1600 (think Nikon's ISO100 is not real).

[dreed]

When you normalised the data to match 5D resolution, did that include adjusting for sensor size for the 7D?
Or was the normalisation just based on megapixels?

Are you interested in doing this to include the Canon S90, which DxO is suggesting actually does quite well?

[Guillermo Luijk]

When you normalised the data to match 5D resolution, did that include adjusting for sensor size for the 7D?
Or was the normalisation just based on megapixels?

Are you interested in doing this to include the Canon S90, which DxO is suggesting actually does quite well?

The normalisation is only based on Mpx, and it must be so. Normalising by the sensor size is wrong unless you want to compare quality per unity of surface of the sensor. This is useless if what we want to compare is the expected quality in final images at a given printed size. The effect of the size of the sensor is already affecting the per-pixel SNR calculations.

If you have the S90 and want to do the shots, go ahead. Remember: tripod, colour chart, uniform lighting, slight camera defocus, several undeerexposed shots and one blown shot, both at ISO100 and ISO1600.

However in the article the complete procedure is explained, you don't need my assistance to obtain the curves.

Regards