[JimmyDranox]

I was looking today on DPR, on image comparison tool. K-1 with Pixel Shift, comparing to other cameras in RAW images. Of course, it was expected that the stacking of 4 image will bring down the noise, but I didn't expect that K-1 PS will destroy any other camera in tems of noise, in any condition, and also, to be on par, or much better in low light/high ISO resolution with 50 Mp cameras, like 645z, or Canon 5DS R. Not even the low pixel FF, like the new Canon 1D X M2, and Nikon D5 are not a match. Only Sony A7S II, with his only 12 Mp, come close.

But switching to JPEG, is not the same story, at least in the DPR tests. Even if the resolution on high ISO is still the best, noise is dependent on the color, many times better, but sometime not. And maybe here Ricoh can work a little more, on that jpeg engine.

[afan137]

I was looking today on DPR, on image comparison tool. K-1 with Pixel Shift, comparing to other cameras in RAW images. Of course, it was expected that the stacking of 4 image will bring down the noise, but I didn't expect that K-1 PS will destroy any other camera in tems of noise, in any condition, and also, to be on par, or much better in low light/high ISO resolution with 50 Mp cameras, like 645z, or Canon 5DS R. Not even the low pixel FF, like the new Canon 1D X M2, and Nikon D5 are not a match. Only Sony A7S II, with his only 12 Mp, come close.

But switching to JPEG, is not the same story, at least in the DPR tests. Even if the resolution on high ISO is still the best, noise is dependent on the color, many times better, but sometime not. And maybe here Ricoh can work a little more, on that jpeg engine.

Yes, and the dynamic range is also impressive. DPR just release another test to compare DR with Nikon D810, and now is written by more competent Richard Buttler rather than Samuel Spencer.

[Simen1]

Mathematically four combined shots should bring down the (random shot) noise by one stop, but in the DPR tests it looks like its slightly better then the math. I don't know how or why, I just think it looks slightly better then K-1 single shot at one ISO stop lower.

[Fogel70]

Mathematically four combined shots should bring down the (random shot) noise by one stop, but in the DPR tests it looks like its slightly better then the math. I don't know how or why, I just think it looks slightly better then K-1 single shot at one ISO stop lower.

Four images would give 4x image information so the improvement is theoretically two stops. Four ISO 400 shots averaged give the same noise performance as one ISO 100 shot.

Just a fun note: A K1 Pixel Shift image contain about the same amount of image data as one FF 6x7 capture. (Four 36 x 24 mm images capture image data the size of 72 x 48 mm vs 67 x 56 mm for 6x7 )

[JensE]

For random noise of a large class of distbutions, the average of N shots will have a similar distribution as the individual values, but narrower by a factor of a square root of N. I.e. statistically, 4 identically placed exposures only give you a factor of sqrt(4)=2 advantage, or 1EV. However, pixel shift also places green where otherwise there are the 'weak' Red/Blue channels so that estimation of luminosity at those most noisy pixels in terms of luminosity is disproportionally improved, leading to a smoother response.

[pathdoc]

For random noise of a large class of distbutions, the average of N shots will have a similar distribution as the individual values, but narrower by a factor of a square root of N. I.e. statistically, 4 identically placed exposures only give you a factor of sqrt(4)=2 advantage, or 1EV. However, pixel shift also places green where otherwise there are the 'weak' Red/Blue channels so that estimation of luminosity at those most noisy pixels in terms of luminosity is disproportionally improved, leading to a smoother response.

Someone in the know needs to write "DSLR Signal Processing for Dummies." I would happily read it, and popcorn futures would hit a new high.

[Fogel70]

For random noise of a large class of distbutions, the average of N shots will have a similar distribution as the individual values, but narrower by a factor of a square root of N. I.e. statistically, 4 identically placed exposures only give you a factor of sqrt(4)=2 advantage, or 1EV.

1EV = sqrt(2)=1.414, sqrt(4)= 2 is 2EV and sqrt(8)= 2,828 is 3EV ....

This is also easy to check. You get the same noise result if averaging 4 images, or use 4 times the exposure.
Capture 4 images using ISO 400 and average them, or capture the same image using ISO 100 with four times the exposure.

[JensE]

Sorry, my conversion of noise amplitude to exposure times was indeed off, thanks for pointing that out.

Taking four times as long of an exposure (time averaging) is equivalent to averaging four exposures of the same exposure time (sample averaging), i.e. the system is at least roughly ergodic. As EV=log2(f^2/t) with f being the aperture number and t the exposure time, a four times longer exposure clearly gives a -2EV difference.

[pathdoc]

Sorry, my conversion of noise amplitude to exposure times was indeed off, thanks for pointing that out.

Taking four times as long of an exposure (time averaging) is equivalent to averaging four exposures of the same exposure time (sample averaging), i.e. the system is at least roughly ergodic. As EV=log2(f^2/t) with f being the aperture number and t the exposure time, a four times longer exposure clearly gives a -2EV difference.

Natural log, or log to base ten? At the moment I am reading this as "square the f stop, divide by exposure time, double it and take the base ten log of the result."

[JensE]

Natural log, or log to base ten? At the moment I am reading this as "square the f stop, divide by exposure time, double it and take the base ten log of the result."

Log to base 2.

[pathdoc]

Log to base 2.

I see; thanks.

IOW log2 (aperture reading squared x number on the shutter knob). Because you're dividing by 1/x of a second, simply multiply by x.

It gets even easier because all of the whole-stop aperture numbers are easy square roots anyway. The hardest part is the log, and even then, exponents of 2 should not be too hard to memorise and extrapolate as required. A quick mental calculation will probably get you in the ballpark most times, and IIRC Sunny Sixteen is EVISO 10015, and EVISO 1000 is f/1 for 1 second at ISO 100.

ETA Wikipedia. EV (any ISO) = EV at ISO 100 + Log2(chosen ISO/100)