Are Digital Sensors There Yet?

[Dj Joofa]

Some of the recent discussion made me think about the resolution of a digital sensor vs. a film scanner, and both compared to the perception of human vision. For a digital sensor or film scanner to be an accurate indicator its resolution should be at least equal to the resolution of human perception.

We shall assume the 1% rule of just noticeable difference (JND) for human perception. We shall not consider the effect of illumination change on the contrast sensitivity of human eye. We shall also assume that the bit depth of the sensor analog to digital converter (ADC) and the number of stops allowed are approximately the same, which corresponds to the behavior of many digital sensors out there in the market. For e.g., the Red One sensor has a current bit depth of 12 bits and reportedly 11 or 12 stops.

For a digital sensor the number (or index) 1 should be assigned to the minimum physical signal the sensor can detect. Now whether the first few bits contain noise, either because of temporal noise or the resolution of quantization, is besides the point. At this stage we are just trying to determine its resolution vis-a-vis human perception of the same signal. Hence, we shall also assume the minimum signal the sensor can detect is also the same darkest black that the eye can see.

With a JND of 1%, the point where the resolution of a digital sensor becomes finer or equal to the human perception may be found by solving the following euqation:

1.01^(x + 1) - (1.01)^x = 1

which gives x = 463.

The above number is closely tied to our assumption that the bit depth of the ADC and number of stops are the same. Otherwise a digital sensor has to increase the number of bits to bring itself under human resolution. For e.g., recall the Red One sensor has currently a 12 bit ADC (and 11-12 stops), and if it wanted to operate at the level of 1% JND resolution, then it shall have to opt for an ADC with a minimum of 19 bits.

This number, 463, is the same one which I mentioned in an earlier post that corresponds to an object that is 100 times more luminous than the darkest black in an image. Lets grab the full significance of this number. In a typical Cineon / DPX scan the darkest black is mapped to 95 and the reference white to 685. Using the standard gamma for density to log of incoming light intensity and the default resolution of a densiometer, calculation gives me that 100 reference white will be mapped to 695. This is about (695 - 95) / 1024 = 58.6% of the range of a 10-bit DI film scan. Hence, for about close to 60% of the "film range" the resolution of the digital sensor is not enough and only gets better afterwards.

Now lets see how does the densiometer for a film DI scan compare? The resolution of a densiometer can be varied, however, the default value used to derive the reference black of 95 and reference-white (close to 90% brighter object) is 0.002. A JND of 1% will be mapped to a density of 0.00259 which is more than the resolution of the densiometer. Hence, film DI scans are good representative of human perception.

Any thoughts? Comments?

[Andrew M.]

And what we do with the fact that human eye can resolve only 6.5 stops?
http://en.wikipedia.org/wiki/Human_eye

[Antoine Fabi]

Hi Andrew,

Nice find.
I can't see why it could be true though...
If i watch outside by the window, and i can still see OK inside, which a camera can't...

Interesting...i am curious...

[Radoslav Karapetkov]

I was just thinking of googling on the latitude of the eye.

20 stops with dark adaptation... not bad...

So, it appears that I'm naturally equipped with a couple of BROWN One 40 K cameras.

The question is - why would I need a RED One then?..

All I have to do is find my LEMO port... gotta be here somewhere.... wait... OUCH! :greedy:

:clown2:

[Petr Dvorak]

karapetkov we all just wait for some wetware interface :)

athough there is one blind guy who has plug back on his head connected with camera glued to his black glasses. He get b/w picture in size aprox 8 x 8 pixel. And guess what he can drive a car with this resolution! That is a miracle. :)

look
http://www.popularmechanics.com/science/health_medicine/1281076.html

[Dj Joofa]

And what we do with the fact that human eye can resolve only 6.5 stops?
http://en.wikipedia.org/wiki/Human_eye

It is difficult for me to believe that human eye is limited to 6.5 stops even in static environments. The window example given by Antoine is a good one.

[Andrew M.]

Eye has good aperture adjustment.
The best example is plasma versus LCD.
LCD has less contrast ratio but most of LCDs have dynamic brightness.
Yes during the day LCD is turned up a bit and at the evening down and this way is usable in the most living rooms.
Plasma has sufficient contrast so don’t have to be much adjusted.
Yet side by side plasma color saturation looks always better then LCD.

Same with the human eye, 20 stops dynamic range but only 6.5 static.

We are easily blinded if we do not have enough time to adjust our “aperture”
Have you ever drove in to the badly lit tunnel from bright noon light?

[Antoine Fabi]

euh...so why am i able to see outdoors and indoors at the SAME time ???

I understand the difference between static and dynamic, but i'm sure i see more than 6.5 f/stops at the same time, unless a video camera has "real" 3 f/stop dynamic range.

[Radoslav Karapetkov]

athough there is one blind guy who has plug back on his head connected with camera glued to his black glasses. He get b/w picture in size aprox 8 x 8 pixel. And guess what he can drive a car with this resolution! That is a miracle. :)

Wow, that's really something.

Science playing Jesus...

Still, reminds of The Matrix somehow.

What if someday a nameless CIA or KGB operative would be able to mess with your brain via Bluetooth or Irridium.... Creepy.

Maybe I read too much George Orwell.

Life is beautiful. :innocent:

[Petr Dvorak]

teleportation is reality, invisibility is very near, 4K vaporware is real too... what next?

[Radoslav Karapetkov]

teleportation is reality

Please continue...? :)

[Alexander Nikishin]

Please continue...? :)

Yes, please do!

[Petr Dvorak]

Its getting pretty scary in our space, huh :shiftyph34r:
they did it in Austria at Innsbruck University back in 2004 on this machine
http://heart-c704.uibk.ac.at/recent/teleportation/quantumteleportationsetup.jpg

They just teleported one particle, I think it was a foton, over 1 meter far.
http://heart-c704.uibk.ac.at/recent/quantumteleportatione.html

... and then they reached 600 meter distance over Danube river!
http://news.bbc.co.uk/2/hi/science/nature/3576594.stm

Beam me up Scotty!

[Dj Joofa]

Hence, we shall also assume the minimum signal the sensor can detect is also the same darkest black that the eye can see.



Interesting I am quoting myself! Anyway. In actuality the eye can see at much lower lux level than the minimum specified for a digital sensor, hence, when I assumed that eye and digital sensor have the same minimum sensation of physical signal, that is actually giving much leeway to a digital sensor, and the actual situation is even worse than my calculations in my quoted message.

In addition, please note that any Gamma curve applied to the linear data output from ADC is not going to rectify the situation a whole lot, because the linear sensor has already crushed several analog levels (which a film scanner can see on a film negative) on a digital sensor into the same bin, hence there is some inherent posterization going on here, with the loss of some local detail.

There are a number of ways to rectify the resolution of a digital sensor.

(1) Increase the sensitivity of the sensor so that its resolution becomes less than the minimum of human eye's, hence, in effect increasing its dynamic range. However, that is not an easy step to do in practice.

(2) Use a non-linear sensor, such as a logarithmic CMOS. There is a lot of literature on this and one can do a Google search to get acquaintance with this subject.

[Paul Leeming]

teleportation is reality...
Hehe, seeing as I'm living ten years in the future here in Tokyo, I can confirm that teleportation is a reality.... we even have a train station which serves as the central locus for it!

http://www.visceralpsyche.com/misc/web_images/tokyo_teleport.jpg

:)

Paul

[Adrian T.]

And what we do with the fact that human eye can resolve only 6.5 stops?
http://en.wikipedia.org/wiki/Human_eye

That's not true. It's between 13 stops and 20 stops.

Read the section "The Dynamic Range of the Eye" in this article: http://www.clarkvision.com/imagedetail/eye-resolution.html (http://www.clarkvision.com/imagedetail/eye-resolution.html)

10,000:1 is about 13 stops
1,000,000:1 is about 20 stops

[Alexander Nikishin]

They just teleported one particle, I think it was a foton, over 1 meter far.
http://heart-c704.uibk.ac.at/recent/quantumteleportatione.html

... and then they reached 600 meter distance over Danube river!
http://news.bbc.co.uk/2/hi/science/nature/3576594.stm

Beam me up Scotty!

I love technology. One of my favorite's is still the Philadelphia Experiment.

http://en.wikipedia.org/wiki/Philadelphia_Experiment

[Steve Sanacore]

Copying a particle and actually moving it are not the same thing. I wonder what they really did and how they can prove it. I would love to know there is hope for such an invention.

As far as our eye's dynamic range - yes I believe it is much greater than 6.5 stops. I always thought it was at least about 12 or more..

I am sitting in the Caribbean right now by a large window. I can see detail in the clouds and still see detail in the shadows at the same time without moving my eyes. I wish I had my spot meter with me, but it has to be at least 10 stops. And yes if I move my focus from the cloud to the shadows, my eye opens up a few stops and I can see more in the shadows but I can still see detail in the clouds. Amazing.

I always dreamed of the day when film or now of course sensors will have this range. We will be able to really capture images with more subtlety, with so much less gear.

[Graeme Nattress]

In addition, please note that any Gamma curve applied to the linear data output from ADC is not going to rectify the situation a whole lot, because the linear sensor has already crushed several analog levels (which a film scanner can see on a film negative) on a digital sensor into the same bin, hence there is some inherent posterization going on here, with the loss of some local detail.

A linear sensor does not crush. It keeps linear. Putting an expansion in the sensor with a log or gamma curve doesn't magically make more photons hit in the shadows to brighten things up a bit. As long as the AtoD quantiaation noise is lower than the noise of the shadows that you're reading, you're not losing a thing. Non-linear sensors are more about recording less light as more light hits, not expanding out the shadows.

Graeme

[Dj Joofa]

A linear sensor does not crush. It keeps linear. Putting an expansion in the sensor with a log or gamma curve doesn't magically make more photons hit in the shadows to brighten things up a bit. As long as the AtoD quantiaation noise is lower than the noise of the shadows that you're reading, you're not losing a thing. Non-linear sensors are more about recording less light as more light hits, not expanding out the shadows.

Graeme

Oh, the linear sensor is going to crush very subtly. Consider this. The minimum physical signal that sensor can sense is by definition index or number 1. Now the next level the signal can see is number 2, which means the physical light intensity doubled. However, our assumption was that the minimum signal seen by the sensor is also the minimum the human eye can see. Now the next level the human eye can see is 1.01, here is how the human eye will assign levels:

1, 1.01, 1,02, 1,03, ..., 2.00. Note that there are 69 levels the eye will see between the physical intensity doubling from 1 to 2. Since our linear sensor can only "see" levels 1 or 2 ad not in between, it will assign these 69 levels to either 1 or 2, hence, by the time the physical intensity doubled, it has already crushed 67 ( = 69 - 2) levels to the either bin 1 or 2. Now raising them to any Gamma is also not going to bring the detail back, which was there between 1 and 2.

[Dj Joofa]

A linear sensor does not crush. It keeps linear. Putting an expansion in the sensor with a log or gamma curve doesn't magically make more photons hit in the shadows to brighten things up a bit. As long as the AtoD quantiaation noise is lower than the noise of the shadows that you're reading, you're not losing a thing. Non-linear sensors are more about recording less light as more light hits, not expanding out the shadows.

Graeme

Putting a log in between does help, not in terms of magical photons, but by expanding the range at the lower end of the spectrum and contracting at the higher end, so that now all levels in the log space have the same distance between them, which previously was not the case, for e.g., levels 2 and 1 have a distance of 1.01 - 1.0 = 0.01 where as the difference between levels 1001 and 1000 is 21168 - 20959 = 209! In the log domain all difference will be same and then linear sampling after applying log can easily see all levels (with some gain if it needs to raise the level of resolution of quantization).

[Andrew M.]

Yes, and this way you save some space on your disk and maybe some CPU
You will not make your picture better.

[Steve Sanacore]

Oh, the linear sensor is going to crush very subtly. Consider this. The minimum physical signal that sensor can sense is by definition index or number 1. Now the next level the signal can see is number 2, which means the physical light intensity doubled. However, our assumption was that the minimum signal seen by the sensor is also the minimum the human eye can see. Now the next level the human eye can see is 1.01, here is how the human eye will assign levels:

1, 1.01, 1,02, 1,03, ..., 2.00. Note that there are 69 levels the eye will see between the physical intensity doubling from 1 to 2. Since our linear sensor can only "see" levels 1 or 2 ad not in between, it will assign these 69 levels to either 1 or 2, hence, by the time the physical intensity doubled, it has already crushed 67 ( = 69 - 2) levels to the either bin 1 or 2. Now raising them to any Gamma is also not going to bring the detail back, which was there between 1 and 2.

Why can't the sensor see a 1% or 10% rise in signal instead of only 100% rise? I don't understand why it has to go from 1 to 2 and double the intensity.

[GlennChan]

We shall assume the 1% rule of just noticeable difference (JND) for human perception.
See Matt Cowan's post here:
http://www.cinematography.net/Pages%20DW/ColorBitDepth.htm

2- For film scans, you actually don't have to scan at a very high bit depth at very high spatial resolutions. The film grains themselves don't carry that much 'bit depth' information. Peter Swinson has pictures of this (I don't believe they're currently up anywhere on the Internet though).

3- In terms of how much bit depth we need, it does depend on:
A- How much noise there is in the system.
B- Characteristic of the source signal. If you are showing very large test patterns (i.e. very large visual angle) with a noise-free source, then yes you will need a lot of bit depth. For most real world images, you won't see banding.
C- Dithering technique used. This however can raise the noise in the picture.

Notice that TV manufacturers can (mostly) get away with 6-bit LCD panels. We just don't see any banding artifacts most of the time.

In terms of image acquisition, both film and video cameras have their own forms of noise so you mostly won't have problems with banding due to insufficient bit depth.

4- Some cameras have banding issues due to insufficient bit depth in (presumably) the signal processing. Example:
http://dvinfo.net/canon/articles/article10.php

I think the F900 suffers from this problem too (though the footage I've seen was dubbed to DVCPRO HD, then DVCPRO50).
Red shouldn't have this problem when the digital signal processing would be done on the GPU (at 24/32-bit floating point precision). *Compression is on-board the camera (where it's possible for the Red team to screw it up). Red also doesn't implement a knee algorithm, so that might be it.

Note that this is a different problem than not having enough bit depth in the sensor, but results in worse looking banding artifacts.

5- In my opinion, people generally pay too much attention to theory and numbers but not real world observations. (Ok ok I do that too. :D )

But insufficient bit depth is generally not an issue. What you should worry about is:
A- Banding / posterization issues in some rare cases with particular cameras. (In a small amount of F900 footage I've seen it's noticeable.)
B- Bit depth / banding issues when working with CG material, since they are free of noise.

C- Other, more important issues. In the case of the F900 footage I saw, the knee algorithm in the camera looked like it was doing evil things to the image (it doesn't look good with very bright, saturated stage lighting).

[Andrew M.]

That's not true. It's between 13 stops and 20 stops.

Read the section "The Dynamic Range of the Eye" in this article: http://www.clarkvision.com/imagedetail/eye-resolution.html (http://www.clarkvision.com/imagedetail/eye-resolution.html)

10,000:1 is about 13 stops
1,000,000:1 is about 20 stops

I read it, guy is mixing up the sensitivity of the eye with the ability to recognize usable contrast between two elements of the picture, kind of MTF of the eye.
Like rating lenses at 200lp/pmm at 10% MTF gives you much higher value then for usable contrast at 50% MTF.
He also talks about recognizing different star brightness, again give it time to adopt and we can have 20 stops.

The thing is, how much contrast level we can detect on the single chart without moving our eyes at all and still recognize if it is letter L or X there.
Remember for digital sensor we have to settle on acceptable quality level.
If we will just talk detection of difference then it is quite a different story.

When you move your camera in one second from the picture of the star at night to the brightly lit room across the street with simultaneous zoom in of FL 800 and adjusting the aperture, and you just need to recognize a difference of one detail, sure you will get 20 stops out of RED Misterium this way.


Remember human eye switches to black and white at low illumination, kind of night vision. We are talking color DR. If you take black and white optimized Misterium you probably can get extra 6dB in DR as well.

With human eye the things are bit more complicated, since peripheral vision has completely different characteristics then middle focus area. Also the ability of night vision of our eye throws the wrench in to the estimation.

I guess, our peripheral area (say 10% off the center line) during the day has very low DR, however giving it time we can adapt to the lighting condition.

The best experiment will be to take a chart and check it, not only for in focus but also for minimum 30% of peripheral area and see what we can differentiate in terms of color light level by reading the large color text on the chart. Remember acceptable noise/picture quality have to be picked beforehand.

[Graeme Nattress]

At the low levels you're talking about, both the eye and sensor are necessarily swamped by noise. At that point, the eye is doing overlapped multi-second exposures to get the noise down also.

Graeme

[GlennChan]

Oh, the linear sensor is going to crush very subtly. Consider this.


1, 1.01, 1,02, 1,03, ..., 2.00. Note that there are 69 levels the eye will see between the physical intensity doubling from 1 to 2.

Your example is misleading because, in a way, you're underexposing.

You'd raise the exposure such that there are 100, 101, 102, 103, etc. units of light hitting the sensor. Therefore quantization noise is not as significant.

[Dj Joofa]

At the low levels you're talking about, both the eye and sensor are necessarily swamped by noise. At that point, the eye is doing overlapped multi-second exposures to get the noise down also.

Graeme

I think I am not getting my point across. Always happens with me. Sad story.

I am not asking you to operate in this low level between 1 and 2. That was just for illustration. The whole point of discussion was with a typical linear sensor, with the assumptions I put in, its resolution will only match the human eye at index 463 -- which is a big index, remember we are talking about very bright objects now -- 100 times brighter than the lowest level so its kind of like early highlights, and also I mentioned that this is about 60% of the range of a typical fllm DI scan. And this is the range where most of the film content is, your actors, dogs, pets, cats, lizards, heroes, ....

This is the bottom line: According to the assumptions I put in, which actually gave much leeway to the digital linear sensor, its resolution will only exceed a human's after only 60% of film range has exhausted, where as the typical film DI scanner working off a film negative will be okay from the very start according to the numbers in my first post.

[Dj Joofa]

Your example is misleading because, in a way, you're underexposing.

You'd raise the exposure such that there are 100, 101, 102, 103, etc. units of light hitting the sensor. Therefore quantization noise is not as significant.

Please see my response to Greame, I am not asking you to underexpose. The whole point is that unless you have exhausted the real usable film range of exposure (about 60%) your linear digital sensor is only giving you exact values at integer boundaries and crushing values in-between according to my assumptions.

[Luis de la Cerda]

Trying to compare human vison to a video camera is an excercise in futility. Both have very different characteristics in how they work. From what I understand, the human eye would make a lousy camera to say the least. It has decent resolution in the fovea (the very center of our vision), but that area is tiny (around 1%) compared to our complete FOV. The fovea is also the only place we have color sensing elements. Everywhere else (99%), our vision is very very low resolution and black and white. It's so low res in fact, that if your fovea is disabled, you cannot read or even recognize faces for that matter. But this large area of our eyes is important because it has much more contrast sensitivity than the fovea does, as well as motion sensitivity, giving the brain valuable information that results in the picture we see. The brain is the real wizard behind human vision, not the eye. Our eyes are moving constantly looking for contrasty areas detected by peripheral vision, scanning in detail with the fovea and storing all the information as a virtual image. The brain reconstructs all this data into the image we see, but most if this image is not real but rather a reconstruction from bits and pieces gathered with different areas in our eyes. This is why a magician can fool our vision when performing magic, or why we perceive a series of still frames as motion (persistance of vision). In fact, there's a condition where the eye's muscles fail and the eye becomes static. In this condition, even though the eye remains fully functional, the subject becomes completely incapacitated and technically blind.

So, to answer the initial question in this thread, sensor technology is there. DSP is what's still lagging way behind.

[GlennChan]

Joofa: Ah, I think I know what you're getting at. Essentially, what you're saying is that current A->Ds have insufficient precision at the dark end of the scale. (Similar to what Charles Poynton calls the code 100 problem.)

I've wondered about that myself.

In practice I don't think it is ever really a problem. I think sensor noise effectively dithers away any banding/posterization artifacts. Perhaps if one were to build an extremely noise-free sensor (the Mysterium-ier sensor?) then this could be an issue??

[Michael Booth]

Some very interesting posts, very educational, thanks guys.

[Dj Joofa]

Trying to compare human vison to a video camera is an excercise in futility. Both have very different characteristics in how they work. From what I understand, the human eye would make a lousy camera to say the least. It has decent resolution in the fovea (the very center of our vision), but that area is tiny (around 1%) compared to our complete FOV. The fovea is also the only place we have color sensing elements. Everywhere else (99%), our vision is very very low resolution and black and white. It's so low res in fact, that if your fovea is disabled, you cannot read or even recognize faces for that matter. But this large area of our eyes is important because it has much more contrast sensitivity than the fovea does, as well as motion sensitivity, giving the brain valuable information that results in the picture we see. The brain is the real wizard behind human vision, not the eye. Our eyes are moving constantly looking for contrasty areas detected by peripheral vision, scanning in detail with the fovea and storing all the information as a virtual image. The brain reconstructs all this data into the image we see, but most if this image is not real but rather a reconstruction from bits and pieces gathered with different areas in our eyes. This is why a magician can fool our vision when performing magic, or why we perceive a series of still frames as motion (persistance of vision). In fact, there's a condition where the eye's muscles fail and the eye becomes static. In this condition, even though the eye remains fully functional, the subject becomes completely incapacitated and technically blind.

So, to answer the initial question in this thread, sensor technology is there. DSP is what's still lagging way behind.

I do not know why people are having hard time understanding what I am proposing. I am not comparing digital sensor with human vision. I am comparing a digital sensor with a film DI scanner and the criterion for discrimination between them is the human vision's resolution.

My basic point is that instead of doing ADC first and then Gamma/Log as done in a typical digital video imaging system, we should follow the model of film and DI scanner and do Log/non-linear first and then ADC (with some gain if needed before ADC) so as not to loose some detail which we keep on loosing until early highlights are encountered.

I guess, enough said.

[Dj Joofa]

Some very interesting posts, very educational, thanks guys.

You are welcome. I think this forum is great and we shall all be grateful to Jim and Co. for providing us a forum for venting out our analog frustrations after digitizing them.

[Dj Joofa]

Joofa: Ah, I think I know what you're getting at. Essentially, what you're saying is that current A->Ds have insufficient precision at the dark end of the scale. (Similar to what Charles Poynton calls the code 100 problem.)

I've wondered about that myself.

In practice I don't think it is ever really a problem. I think sensor noise effectively dithers away any banding/posterization artifacts. Perhaps if one were to build an extremely noise-free sensor (the Mysterium-ier sensor?) then this could be an issue??

Exactly. However, the problem manifests itself not just at the dark extremes but all the way to early highlights only with less and less impact. Now this would be an empirical study involving humans that where is the impact really lost.

You are also right about sensor noise. The sensor noise acts as "dithering" mechanism (similar to random noise added as dither to audio *before* ADC to reduce the perception of harmonic distortion resulting from quantization), and actually helps reduce the posterization in an image. This is a very significant point.

[GlennChan]

So, to answer the initial question in this thread, sensor technology is there. DSP is what's still lagging way behind.
In my opinion (I'm just guessing here, as is everyone else), the way the human eye works is fundamentally different than how a video camera works. The human eye has to be constantly moving... as you point out, a non-moving eye doesn't work very well.

A moving sensor allows you to have resolution exceeding the Nyquist limit and good aliasing characteristics. Whereas a fixed sensor is limited to resolution no higher than the Nyquist limit, with tradeoffs between resolution and aliasing.
The Arriscan or any flatbed scanner would be an example of how moving sensors can be a good design.

2- For digital cinema, I think the limitation is in the sensor.

With the Red, the DSP is not that big a deal because it can be non-real-time (e.g. image processing is done in Redcine, Red Alert).

I think the real limitation right now is in the sensor (more dynamic range, less noise, higher frame rates, less aliasing + other artifacts, etc. etc. would be desirable traits).

3- In some aspects, our vision has extremely high performance.

Vernier resolution gets up to something like 600cycles/degree in lab situations. (Far beyond 4K.)
Dynamic range well beyond what's achievable in a theatre.
Extreme low light sensitivity (though our vision becomes black and white).
Stereo vision (ok, it doesn't work that well and people with lazy eye aren't that affected by the lack of stereo vision).
etc. etc.

[GlennChan]

My basic point is that instead of doing ADC first and then Gamma/Log as done in a typical digital video imaging system, we should follow the model of film and DI scanner and do Log/non-linear first and then ADC (with some gain if needed before ADC) so as not to loose some detail which we keep on loosing until early highlights are encountered.
It might be that film scanners have a linear A->D, and then some gamma/log/whatever curve applied afterwards. (I don't design them but I believe that's how they are designed.)

[Dj Joofa]

It might be that film scanners have a linear A->D, and then some gamma/log/whatever curve applied afterwards. (I don't design them but I believe that's how they are designed.)

I don't think so. Because they don't need to. The non-linear log has already been incorporated as density change on the film negative. Hence, all a film DI scanner has to do is to linearly digitize the non-linear log space coming from film negative density.

[Dj Joofa]

Why can't the sensor see a 1% or 10% rise in signal instead of only 100% rise? I don't understand why it has to go from 1 to 2 and double the intensity.

Because by definition the index 1 is already the lowest level of physical sensation it can detect hence all other indices are actually (integer) multiples of this lowest physical value.

[Graeme Nattress]

As long as the bit depth of the AtoD is such that it's quantisation noise is lower than that of the noise in the signal you're sampling, then you're getting it all. By putting an analogue curve ahead of the AtoD and assuming that amplifier adds no noise, you're just wasting bits on noise. Now, I've seen it in some cameras that that analogue noise may look nicer than the quantisation noise, but it's still noise of similar magnitude and still of the same magnitude.

The Cineon file will represent 3.3 scene OD as a 2.0 neg OD, as code values such that each code value represents 0.002 OD on the neg. But there's compression curves at either end of the linear response section, so you get more like 4.0 scene OD on a very good neg, if you're luck, but with decreasing fidelity at the top and tail.

Graeme

[Bing Bailey]

Honestly I think at this point we're not gaining much when it comes to resolution. yeah we could go to 8k but its really the contrast and colour performance that will make the difference. 4k gets us to 35mm film resolution in every real sense. with the right lenses RED ONE is pin sharp. so we don't really need anything sharper. the tools to make sure we get the best of the camera will have more positive results when it comes to using the technology.

if we can obtain the same maximum lattitude as film. the same colour characteristics and contrast we'll be sitting pretty. people do get hung up on lattitude. yeah film has a maximum potential of 15 stops. but you have to look at the word potential. it doesn't mean everything shot on film gets 15 stops of lattitude. although its nice to strive for that. I feel RED ONE is the technology that gets us to about 98% of whats good about film and completely obliterates whats a pain about working with film. RED ONE also offers something that film has always offered and other digital cameras have not really offered. the ability to improve with upgrades in film that's new stocks used on the same cameras. On the Red One thats a firmware upgrade. maybe we should give the builds stock numbers instead :)

[Dj Joofa]

The Cineon file will represent 3.3 scene OD as a 2.0 neg OD, as code values such that each code value represents 0.002 OD on the neg. But there's compression curves at either end of the linear response section, so you get more like 4.0 scene OD on a very good neg, if you're luck, but with decreasing fidelity at the top and tail.

Graeme

Sure, and I mentioned that in my first post in this thread that a 1% JND will map to a density of 0.00259 and hence, the default resolution of the densiometer of 0.002 is finer than that; therefore, from the get-go the film DI scanner has advantages over a typical digital sensor which shall take 60% "film range" to get to that point.

[Graeme Nattress]

The best guide to this issue, as Glenn kindly points out is page 12 of Poynton. As far as I can see from reading it, the 1% JND refers to change in luminance levels at the eye.

Now, follow that through on the film scan side, code values are 0.002 neg OG, which is 0.002 / 0.6 = 0.0033 scene OD (assuming we're in the linear region of film with a film gamma of 0.6). Now that represents a change of 0.0033 / 0.3 stops or 0.011stops change in luminance, or a multiplyer of 1.007. That's just below our 1% JND. As film gamma or slope of the transfer get more shallow in the shadows and highlights, it will surely move beyond the 1% JND?

Graeme

[Dj Joofa]

Now, follow that through on the film scan side, code values are 0.002 neg OG, which is 0.002 / 0.6 = 0.0033 scene OD (assuming we're in the linear region of film with a film gamma of 0.6). Now that represents a change of 0.0033 / 0.3 stops or 0.011stops change in luminance, or a multiplyer of 1.007. That's just below our 1% JND. As film gamma or slope of the transfer get more shallow in the shadows and highlights, it will surely move beyond the 1% JND?

Graeme

Thanks for all your effort in going through Poynton and responding to me with numbers.

Yes, I agree with you regarding some portions of shadows and highlights. We may fall outside 1% JND in these small regions. However, for most parts we shall stay within 1%. Also, though you derive the multiplier as 1.007, I derived the factor as 1.3, and here is how I got it (though my calculation is just the opposite of yours):

log_10 (1.01) * 0.6 = 0.00259 density on negative. And,

0.00259 / 0.002 = 1.3

Thanks again for taking the effort.

[Graeme Nattress]

I don't have the curve diagram handy, but the amount of values that fall in the non-linear region are significant, rather than small. The other issue is that noise and grain severely complicates the pure math of it all, in that if we were grainless and noiseless, we'd have a much greater issue that we do.

Graeme

[Radoslav Karapetkov]

OMG, this thread grew beyond me. :)

[Alexander Nikishin]

OMG, this thread grew beyond me. :)

Isn't it great!

Still fun to read and attempt to understand such specialized topics.

Wikipedia helps a bit.

[GlennChan]

I believe if you google "code 100 problem", you'll get an excerpt from Charles Poynton's book. Though you still have to understand what poynton is saying.

2- Perhaps an easier way of understanding things is to play around in Photoshop.

Take a gradient.
Duplicate it, and add noise into that version.
Add two levels adjustment layers.
On one of the levels, make the mid/gamma 2.222
On the other one, put in 0.45
Make sure you are in 8-bit RGB mode.
Toggle the noisy layer on-off, and toggle the levels pair on/off. Note that the noise does a very good job of hiding the banding artifacts.

You can also play with the dither setting in the gradient generator.


The non-linear log has already been incorporated as density change on the film negative. Hence, all a film DI scanner has to do is to linearly digitize the non-linear log space coming from film negative density.
I don't think so. If you wanted to undo the non-linear response of film (which nobody does in practice), then that would make quantization errors even worse.

If you left the film's response alone, you have the same problem as with video cameras (except that the film scanner probably implements a 16-bit A->D instead of 12-bit/14-bit/whatever).

[Dj Joofa]

I believe if you google "code 100 problem", you'll get an excerpt from Charles Poynton's book. Though you still have to understand what poynton is saying.



I have Poynton's book -- it's a great book.



2- Perhaps an easier way of understanding things is to play around in Photoshop.


I don't think so. If you wanted to undo the non-linear response of film (which nobody does in practice), then that would make quantization errors even worse.

If you left the film's response alone, you have the same problem as with video cameras (except that the film scanner probably implements a 16-bit A->D instead of 12-bit/14-bit/whatever).

I do agree that noise can hide the imperfection of a linear digital sensor in the most useful range. However please consider what I write below:

If you undo a DI film scan you shall get all levels back in the most useful range that I define below -- because nothing was lost in the 0.6 gamma range which according to Arri's DI manual extends from close to 2% black to well above 100% bright range. The reason is that as I explained in a previous post, the 1% JND mapped to 0.00259 density on film negative in this range, and a typical DI scanner has a resolution of 0.002. Film negative is an analog medium; therefore all information is there, our sampling resolution (0.002) is finer than an increase in one perceptual level that a human eye can see (0.00259 density), and hence, we are fine.

This is the most useful range of film content -- about 60% of a typical DPX scan lies in this range -- hence, a lot of our film content such as actors, houses, and dinosaurs are in this range. And this is the range where film scanner has finer resolution than human eye and a linear digital sensor has poorer.

Tom Cruise was most probably exposed within this 2% to 100% range. In this range the linear digital sensor has crushed (463 - 100) / 463 = 78% of detail, i.e., close to 80% detail has been lost -- may be that is why we get that "video look". Where as the film scanner has captured all of it without loosing it.

The remaining 20% is salvaged by using the Gamma curve. Hence, applying the Gamma after the linear ADC is only accurate at integer boundaries and in-between it is blind and thanks to camera temporal noise it gets away to some extent by hiding it.


After this range the linear digital sensor has finer resolution than human vision, however, that is the range in which you will see very bright stuff such as that light bulb, that shinny lamp, the sun in the corner of the beach scene, etc. Julia Roberts was probably not exposed for this range.

Beginning photographers are taught to be mindful of 18% gray as the signal is changing so rapidly in the perceptual space, but the linear digital sensor, and then a Gamma applied afterwards, just saw the true picture at only integer boundaries.

[Dj Joofa]

I also suggest an experiment that shall make you understand how temporal noise is hiding the situation, and one of the few places where noise is actually our friend and we do not want a sensor without one.

I do not know if you will be able to do it but I am sure the people at Red One can do it. This experiment assumes that the internal CMOS chip in the camera is kept at minimum analog gain *before* ADC. If the CMOS is operated at lowest analog gain always and it still gives correct exposure, then please put an ND filter in front of the camera and put that analog gain before ADC to the minimum. Now the resulting image shall get darker. However, *after* ADC, provide gain (I assume that is possible in the chip design) to compensate and bring back the normal exposure of an image and brighten it up to normal. Suddenly, the image might look posterized.

This is because a high analog gain before ADC increases temporal noise and that fills in the quantization gap and provides "dithering" to reduce posterization. This is similar to the situation when dithering random noise is added to audio *before* ADC to reduce the perception of harmonic distortion resulting from quantization.

[Radoslav Karapetkov]

Someone said something about gaining latitude by scaling.

2 stops from 4K to 2K, for example or something like that.

The post was here or in another similar thread.

Can anyone elaborate on this, sounds very interesting?

Or just give a link if it has been discussed before.

Thanx.

[Antoine Fabi]

Isn't it great!

Still fun to read and attempt to understand such specialized topics.

Wikipedia helps a bit.

heh heh...I can understand the principle, the logic, but the rest leads me to humility and faith. :) :) :)

[Dj Joofa]

I think I shall also elaborate one more point that I brushed upon in an earlier post of mine when I said that one solution to this problem of a linear digital sensor not having enough resolution is to have more dynamic range.

Lets pick a real case for Red One camera. We have a number of options here.

(1) Increasing bit depth. My understanding is that Red Camera ADC is 12-bit, and if in linear domain Red wanted to keep the same number of stops as currently it has, and also wanted to have enough resolution (matching human eye) then Red would have to opt for an ADC with at least 19 bits. Since, currently, this is not the case, therefore a 12-bit ADC shall crush various perceptual domain signal levels, which a film scanner would be able to recover.

(2) Keep at 12-bit, and increasing analog gain *before* ADC. However, that would mean increasing noise in an image, and losing on actual stops that an image can have from the darkest to the brightest, and may render a high-profile camera to a one with less number of stops.

[Graeme Nattress]

But Joofa, as I said earlier as long as quantisation noise is lower than sensor noise, you're fine. Your point is academic. It's an issue to be noted and why to avoid linear when going down to 10bit or 8bit, but it's not an issue at the sensor because the AtoD is matched to sensor noise in it's resolution. That is why as sensor noise gets lower, you put more resolution on the AtoD.

And no, a film scanner will not help as at the levels we're talking about we're noise limited way before we get to the non-issue you're concerned about.

Basically, the resolution is always noise limited. Higher resolution is meaningless in avoiding banding / posterisation because of the noise. Sampling noise with a finer quantisation is just making more accurate noise, and not increasing resolution.

Now, if you had a totally noise free image, we'd have the problem, but that's an issue for CGI, not film or digital capture.

Graeme

[Dj Joofa]

But Joofa, as I said earlier as long as quantisation noise is lower than sensor noise, you're fine. Your point is academic. It's an issue to be noted and why to avoid linear when going down to 10bit or 8bit, but it's not an issue at the sensor because the AtoD is matched to sensor noise in it's resolution. That is why as sensor noise gets lower, you put more resolution on the AtoD.

Graeme

True. I am in agreement with you here. My sole point of mentioning the 19-bit ADC vs. 12-bit ADC, which as I understand Red is currently using, was more theoretical, as I do not know what noise level Red is encountering in their custom-designed chip. I am sure Red would have wanted to go over 12-bit, but the noise-level in the CMOS would have restricted them. May be next generations of CMOS would address the noise issue.

Thanks for your responses. I have learned a lot from them.

[Graeme Nattress]

I think we will see that as sensors improve, all companies will use better AtoD to go with them. It's inevitable.

Graeme