by Keith Mullin
In this article I hope to explain on a conceptual level how Dual Base ISO works and how Shot Noise can mislead us into thinking it’s just a gimmick from the camera manufacturers. I can assure you right now that Dual Base ISO is not a gimmick of any kind, and by the end of the article you will have a better understanding of what is happening inside the camera and in your images.
If you’ve ever worked with a camera with Dual Base ISO, or just done some research on it, you already know that no matter how many times the manufacturers claim that both ISO ratings have the same or very similar noise profiles, the images produced at higher base ISO settings have a noticeably higher amount of noise. My first experience with this was shortly after the release of the Panasonic EVA1. I was testing it against the Sony FS5 and FS7 cameras and shot it at both its base ISO settings. I was really surprised at how much nosier the higher base ISO setting was. Some of this was a result of the early firmware not having a robust enough codec to really handle the recording and has been much improved with later updates, but much of it was actually due to Shot Noise.
What is Shot Noise?
Shot Noise, at its most fundamental level, is an expression of statistical variance due to the randomness of a subject that is being measured or quantified. References to Shot Noise can be found primarily in electronics, although the principles remain true for any set of measurements, whether in optics, mathematics, or statistics. The concept was first introduced by Walter Schottkey in 1918 in his study of the fluctuation of electrical current in vacuum tubes.
If you are familiar with statistical analysis, you know that the larger the sample size is, the greater the likelihood that the observed results of whatever it is you are trying to measure will be closer to the actual average. With smaller sample sizes, you have a higher likelihood of getting more results that fall outside the normal or expected range. The results that fall outside normal would be the Shot Noise in that set of data.
For example, if you flipped a coin 100 times in a perfect world you would get 50 heads and 50 tails. Since there is a certain amount of randomness in the world the odds that you would actually get equal numbers of heads and tails is relatively small. Multiply that out over many sets of 100 coin flips and you would expect to see that each result falls within a certain range around the actual median result. In statistics we would define a range above and below the average as “standard deviation”, results that fall outside the standard deviation would be statistically anomalous and would be the “noise” described by Shot Noise. In statistics, if too many results fall outside the standard deviation, the results of the analysis cannot be said to be “statistically significant” meaning the overall result can’t be relied upon to give us credible information. In imaging, Shot Noise creates a different result that we will get to later in this article.
To further flush out the theory, before looking at imaging Shot Noise, lets look at an example set of numbers. I used an online coin flipping application to flip a coin 1000 times. I counted (or rather I had it tell me) how many times the coin came up heads. I repeated this 10 times.
|Number of heads||Numerical Variance +/-500||Percentage Variance|
If we declare our “standard deviation” as +/- 2.5% we can see that of the ten sets of 1000 coin flips above, only one of them falls outside the range we have determined is “normal” or “expected”.
In practice we would actually determine our standard deviation by looking at the results and establishing how wide a range is needed to include a majority of results. For our theoretical example we can arbitrarily set it.
Now lets see what happens when we shrink our sample size to just 200 coin flips. Note this is not a randomly selected number, it will have bearing later in this article.
|Number of heads||Numerical Variance +/- 100||Percentage Variance|
As we can see, only three of the results fell within our standard deviation of +/- 2.5%. That is a pretty big change from the first set of data. If we were doing a statistical analysis the second set of results would no doubt be judged to be statistically insignificant since so many results fell outside the standard deviation.
How does this apply to film?
Now that I’ve bored you with all this theory and statistical jargon, let’s look at how this works in digital film.
The concept of Shot Noise was first brought to my attention during a discussion about dual base ISO in digital cinema cameras. This is a relatively new development in digital cinema cameras, and there are only a handful that have dual native ISO. Among them are Sony VENICE, Panasonic Varicam35, Panasonic VaricamLT, Panasonic EVA1, and the Blackmagic Pocket Cinema Camera 4k.
Theoretically a camera with two base ISO’s has the same amount of sensor/gain noise regardless of which base ISO is selected. But in practice every image is noisier at the higher base ISO value than at the lower one. This is Shot Noise.
To understand how we end up with Shot Noise on the higher ISO setting in a dual native ISO camera we need to know a little bit about how dual native ISO is achieved.
Each manufacturer explains the process a little differently, and never in specifics about their own technology. This is understandable because they want to protect their special sauce from being copied by other companies. Regardless of how it is achieved, the principle underneath each is essentially the same.
In a traditional sensor architecture the circuitry is designed so that each photosite has a single capacitor behind it. You can think of the capacitor as a bucket collecting rainwater and the light hitting the sensor as the rain. If the bucket is empty the visual rendering would be black, if the bucket is full the visual rendering would be white. In between full and empty are different shades of grey. (Color information is collected in the same way, but for simplicity in this example we will deal only with the grey scale)
The size of the capacitor or bucket affects how much light is required to fill it up, thus setting the base iso of the sensor. Big bucket, low ISO. Small bucket, high ISO.
In a dual native system there is a second circuit, or bucket, for each photosite. To get the second base ISO, the camera does one of a couple of things, depending on which manufacturer the camera is from. I have heard it explained that in some camera systems the sensor swaps from one capacitor to another of a different size.
And in others that for one base ISO setting both capacitors are used, and when the base ISO is changed will only use one of the capacitors.
Either way, the total capacity for each photosite to read and store light information changes. All of this occurs before the camera does any other processing such as adding gain, thus the two base ISO ratings.
How does this result in Shot Noise?
Think about the capacitors or buckets as one of the samples in our coin toss examples above. Let’s use the Sony VENICE for our example. The VENICE has a low base ISO of 500, this would be the 1000 coin toss sample size. The camera’s high base ISO is 2500, this is the 200 coin toss sample size. Remember that for every stop of sensitivity we gain we halve the amount of light coming in to achieve the same exposure. Since there is a shift of 2 1/3 stops between 500 and 2500 we divide the sample size in half (500 flips), then in half again (250 flips), and then subtract another third of a stop (200 flips).
Lets see what happens, visually, as we change between the two base ISO’s.
Below is an 8×8 grid of light coming off an object in the field of view of the sensor. Each square corresponds to a photosite on the sensor.
Imagine that the light hitting the sensor is grey and of a uniform intensity. We also know that this light has small random variances in the way it hits the sensor and is read. These variances are expressed by the coin toss. If this happened in a world of exacting probabilities, you would end up with an even division of heads and tails each time. But we live in a world of randomness and chaos, so each square in the image below has the number of times heads came up out of 1000 coin tosses for that square, and then the numerical variance above or below the perfect 500 average.
Now lets change +/- count to percentage variance and shift the color of the corresponding squares lighter or darker depending on how far away from the statistical average each square is. We will once again be using our arbitrary standard deviation of +/- 2.5%. If the value of the square is +2.5% or higher the square gets lighter. If the value is -2.5% or lower, the square gets darker. For every additional 2.5% of variance the square would get lighter or darker still, reflecting the greater distance from the average.
Now let’s do this again, but this time at the higher base ISO with a sample size of 200 coin flips.
Let’s see those percentages and the shifts away from the average.
So even though we are not introducing any gain or sensor noise, the light itself creates pixels that register as lighter or darker, this is the Shot Noise in our image.
So how is this better than just gaining up the sensor?
What is interesting is that when gaining up, the Shot Noise behaves the same way no matter how much we gain up. Let’s take the example using 200 coin flips and multiply the results by 5, in essence “gaining up” to the 1000 coin flip sample size.
Once we convert these to percentage variances we can see that the result is exactly the same as before the multiplier was added. This is because the variance is the sample itself, or the light itself in an image.
If all gain did was amplify the signal to fit the larger sample there would be no advantage to having a dual base ISO. However there is something that happens when you add gain to the signal coming off the sensor. You also end up multiplying any errors or artificial spikes from the sensor itself. These appear as colored spots or “snow” in the image. To see this visually, all you have to do is cap the lens of your camera and crank up the gain.
I did exactly this with our Sony VENICE camera and recorded the results in 4k XAVC-I. I was only able to actually gain the lower base ISO up to 2000, as it does not go higher than that, but the results are still quite telling.
You may need to click on the image above to view at full resolution to see the difference between the samples.
In essence, when you gain up you end up with both Shot Noise and sensor noise layered on top of it. Having a dual base ISO camera lets you eliminate one of the types of noise that can appear in your image leaving you with cleaner, more organic looking footage.
Unfortunately there isn’t anything that we can do about Shot Noise. It is purely a matter of physics and would exist no matter how amazing, or how large, or how sensitive the sensor of any camera is. The good news is that with the advent of Dual Base ISO technology the only noise we end up dealing with in lower light environments is Shot Noise.
With these newer cameras, we now have a lot more control over our images and how they look in a wider variety of lighting situations. We no longer have to use multiple cameras if we have need the flexibility to shoot in bright daylight and low light, we can pick just one system and stick with it, making logistics and post much more manageable.