EV and F Stop confusion

A

Andor

Active member
On DxO, Epic Dragon was rated as 14.8 EV. The GH4 is rated at 12.8 Evs.

I read somewhere that 1 EV = 1 ƒ Stop. I don't really believe this yet, because literally everything I'm reading is reporting Dragon as being beyond 14.8 ƒ Stops.

Plus, 12.8 ƒ stops is more than the BMC cameras are reporting, and I just don't think the GH4 will have that much dynamic range.
 
At this point, I'm sort of not of the mind to believe anything any manufacturer says about anything... well... I've been that way for a long time...

But my recent survey of the DSLR manufactuerers that showed at NAB, and whose booth I visited, and asked about 'how their camera's ISO numbers were obtained', did not leave me with the impression that anyone knew other than what the marketing brochure indicated.

So, here is something of the reality...

For an 8 bit representation of image intensity, one only gets 8 'stops' of 'linear' representation... That is 'modified' by using a 'logarithmic' representation, or a piece-wise linear representation on various ranges. So, one could have an 'S-curve-ish' shape, or the like to have less bits representing low or high values, but somewhat linear representation in the middle...

One can extrapolate for deeper bit densities, such as 10-bit, or 12-bit representations.
 
1 EV = 1 stop, that's for sure

Now, measuring DR is not easy, or standardized: you have to decide what counts as meaningful signal and what is just noise. Your choice will greatly affect the numbers: a couple of years back I was measuring "absolute DR" (if there's some signal, it counts, no matter how much noise there is), now I look at "usable DR" (with a very subjective measure of "usable"); I've gone from saying the 5D2+ProlostFlat has 10 stops of DR to saying it's just 6. So: you can't compare sources.

And then, if you just look at DxOmark numbers: what they say is that the Dragon has just 2.5 stops more DR than the GH4... in RAW stills mode. Unless the cinelike profile is better than it looks at first sight, that's probably going to be a 5-stops difference in video mode (my guess).
 
Samuel H said:
1 EV = 1 stop, that's for sure

Now, measuring DR is not easy, or standardized: you have to decide what counts as meaningful signal and what is just noise. Your choice will greatly affect the numbers: a couple of years back I was measuring "absolute DR" (if there's some signal, it counts, no matter how much noise there is), now I look at "usable DR" (with a very subjective measure of "usable"); I've gone from saying the 5D2+ProlostFlat has 10 stops of DR to saying it's just 6. So: you can't compare sources.

And then, if you just look at DxOmark numbers: what they say is that the Dragon has just 2.5 stops more DR than the GH4... in RAW stills mode. Unless the cinelike profile is better than it looks at first sight, that's probably going to be a 5-stops difference in video mode (my guess).
Click to expand...

Here's some pictures from my recent tests for my GH-1. I set the camera at IS0 100, my L-308DC meter at ISO 250, and used the resulting 'F/8 @ 1/50' for my 'meter reading'.

I then varied the F-stop from F/4-F/22 to indicate the 'dynamic range'. That's 6 stops. I think I could go to 6.5(or so), but my lens didn't open wider than F/4...

13926600054_57cc83fddf.jpg


As the histograms indicate I have a little more 'head room' on the F/4 image, perhaps 1/2 stop's worth...

Here's the IRE values from Premiere. As the 'dark' image shows the Waveform indicates there is 'data' in the image, but from the actual image displayed, the grey card is barely visible... and so anything lower than that, is even less visible... which is to say... not useful in an 'effective' dynamic range rating.


13903013251_ca0fb46a64.jpg
 
That's what I changed when I moved from measuring "absolute DR" (is there a signal difference in the waveform?) to "usable DR" (can I see it, and with decent detail, after pushing the image in post?).


But I would change shutter speed rather than aperture: if you can do it without introducing flicker, it is more accurate (in photography lenses, those blades don't always move exactly as much as they're supposed to).
 
I appreciate all the replies! Samuel, I remember you did something on Rolling Shutters once, and I found that super helpful (though I can't find it anymore).

Okay, so now I want to dive deep into dynamic range and... bits... Is it safe to say that 8 bits can only really display 8 stops? And where do displays factor into this (how can I decide the dynamic range of my screen)?

BMPCC is supposed to be 16 bits packed into 12 that then unpack into 16 in Resolve (when shooting RAW). However, I thiiiink ProRes is just a 10 bit codec, so does that mean that despite the 13 stops claimed the camera can do, you'll only really get 10 stops or so if you shoot ProRes?

Why do people use DNxHD (8 bit codec) when there's such a big difference when going to 10 bit / 12 bit / 14 / 16 / etc?

I'm trying to ask the right questions for clear answers, even if the question sounds too constricted, so my apologies.
 
Rolling shutter

8 bits can display 8 stops... if you record linear, which usually only happens if you're recording uncompressed RAW. If you record log, or with a video gamma curve, etc, in theory you can put up to 256 stops in 8-bit. Of course, you'll have one value per stop and it will look like crap, but my point is: the theoretical limit is irrelevant, what matters is if you're willing to work with, say, just 20 values per stop of light. The C100 has way more than 8 stops of DR and it records 8-bit (which looks great).

If you're going to look into bit depth, first you need to look at linear vs log vs video-gamma.
 
Gamma can be confusing. First, all camera image formats are written out as gamma encoded files. The first stage AD in camera may be linear, but even 12 bit raw data files are gamma 1 encoded, I.E they are written with equal bit values assigned to each stop of DR from black clip to white clip. In 8 bit video that is 23 8 bit words for each 10 IRE increment of waveform scale from 0 to 109 with each word being an equal gray scale value change over a nominal 2.2 gamma linear scale.
Gamma 1 is linear voltage value encoding at 3dB per f/stop of light change or 10log(v1/v2). For 12 bit raw that produces a resolution of 315 discrete 12 bit word grayscale values per doubling of voltage output from the sensor over say 13 stops of DR.
Human visual response corresponds to a power density change of 6dB per f/stop or a gamma 2 scale, 20log(v1/v2). Display gamma standards all fall in a narrow range around gamma 2 values to closely match human visual response, but all produce a linear value scale of equal value change per step of increase in input signal. Log gammas introduce curved value distributions on top of the nominal linear display responses to redistribute signal values for aesthetic purposes, typically to represent a broader capture DR in a narrower DR display space.
But all photographic systems, whether film or electronic are designed to reproduce a log gamma response to light stimulus that is visually meaningful to humans. To fully resolve grayscale values to the limits of human visual discrimination requires at least 120 discrete grayscale steps per photometric stop of light values from deep shadow, around 30% on a waveform monitor to around 80%. Human visual discrimination declines rapidly in bright highlight ranges and more slowly in deep shadow, but falls off quite a bit at both ends of the range. Gamma corrected encoding of digital images makes more efficient use of the bit values by not wasting resolution at the top of the scale where it is visually meaningless, and preserving value resolution into deep shadows where it is.
 
Razz16mm said:
Gamma corrected encoding of digital images makes more efficient use of the bit values by not wasting resolution at the top of the scale where it is visually meaningless, and preserving value resolution into deep shadows where it is.
Click to expand...

I wish the various 'low end' manufacturers would publish their 'curves' for their products, rather than having the user 'figure it out'...
 
And add to all of that the constant confusion between Dynamic Range and Latitude. People still keep raving about Alexa's superior DR compared to every other camera out there - but at 14 (and a bit) stops, it's total DR is almost matched by most of its major competitors - so what gives? The Alexa's advantage lies less in it's total DR, and more in its overexposure latitude - in the range it can capture above middle grey.

And at 7 stops over (at it's 800 ISO base), it still has the better part of a stop and half (or more) of highlight room over the F65, Epic, C500, F55 et al. and that makes all the difference.
 
Razz16mm said:
Gamma can be confusing. First, all camera image formats are written out as gamma encoded files. The first stage AD in camera may be linear, but even 12 bit raw data files are gamma 1 encoded
Click to expand...
I think there are several cameras that record linear RAW, not gamma-encoded. I haven't opened the files to look at this, but I'd be almost sure that's the case with the 5D3+RAW, and would think the original BMC probably works this way as well. It's very inefficient, but it doesn't require any processing. The software you open the file with is the one that applies the gamma curve.

I would tend to think the dbolex works that way too, and the si-2k, and any other "dump the sensor data into an uncompressed RAW file" cameras, but you definitely know a lot more than me about those :)
 
Samuel H said:
I think there are several cameras that record linear RAW, not gamma-encoded. I haven't opened the files to look at this, but I'd be almost sure that's the case with the 5D3+RAW, and would think the original BMC probably works this way as well. It's very inefficient, but it doesn't require any processing. The software you open the file with is the one that applies the gamma curve.

I would tend to think the dbolex works that way too, and the si-2k, and any other "dump the sensor data into an uncompressed RAW file" cameras, but you definitely know a lot more than me about those :)
Click to expand...

I don't think so. If it wrote out a linear 12 bit linear file, you would have very poor grayscale resolution below 6 stops down. Each stop would have half the number of grayscale steps of the next higher stop. You could not resolve 13 stops of DR, or even 12 with linear 12 bit encoding. I did a grayscale excel table for 16 bit linear compared to 12 bit gamma encoded for the D16, and compared the results to the number of discrete grayscale levels I could read off a non gamma corrected neutral 12 bit DNG image in Raw Therapee between 20% and 30% value range. It far exceeded what a 16 bit linear file could resolve, much less 12 bit.
 
I think the 5D3+RAW writes 14-bit linear, and the BMC writes 16-bit linear (the result of mixing two 12-bit signals, from the high-gain and low-gain paths). Those should be good enough for the amount of DR that each camera has.

Although arbitrary LUTs can be implemented, the DNG spec uses linear encoding as default.

(yes, I'm bored, and the a7s price doesn't show up no matter how often I refresh the sonyalpharumors or dvxuser browser tabs)
 
You must log in or register to reply here.
Back
Top