AVS Forum banner

All About Gamma

12K views 58 replies 19 participants last post by  fluxo 
#1 ·

Joel Silver, founder and president of the Imaging Science Foundation, explains what gamma is and how it affects the image on display devices, how it remained unstandardized and approximate until 2011, getting manufacturers and content creators to implement the new standard, the importance of setting the gamma control of a display depending on the room lighting, high-dynamic-range proposals from Dolby, Technicolor, Philips, and the BBC and how they might benefit from something other than gamma, the next-generation DRM (digital-rights management) system that is necessary if studios are going to make 4K/UHD content available, answers to chat-room questions, and more.

 

 
 
#5 ·
No mention of how BT1886 response "changes" with respect to black-level measurements? I thought the discussion was supposed to get "uber-geeky."



In other news, so now that we actually have a "gamma" standard ... for the first time ever ... we're going to go ahead and change *everything* again? I'm not sure, but this may be the reason the phrase "going postal" was invented. .... (insert scream here ... you know the one.)
 
#6 ·
I also thought it was strange that Joel presented the information that BT.1886 is not suitable for a bright room, only for a dark room. BT.1886 can compensate for an elevated non-zero black level. If you measured and calibrated to BT.1886 in a bright room, wouldn't BT.1886 be still correct? I do realize the bright room BT.1886 curve will not look like the BT.1886 curve if you measured and calibrated in a dark room.


I'd also like to know what new LCD TV he has in his other room with 550 LEDs since he implies it has that many zones.
 
#9 ·

Quote:
Originally Posted by Stereodude  /t/1533867/all-about-gamma#post_24779165


I also thought it was strange that Joel presented the information that BT.1886 is not suitable for a bright room, only for a dark room. BT.1886 can compensate for an elevated non-zero black level. If you measured and calibrated to BT.1886 in a bright room, wouldn't BT.1886 be still correct? I do realize the bright room BT.1886 curve will not look like the BT.1886 curve if you measured and calibrated in a dark room..

Also curious is how one could have a BT1886 "preset" without knowing where the display's black level is set ... humm ....


Edit: Well I guess you could dial in a "preset" at the factory, and as long as no one changes the black-level/brightness control or white-level/contrast control and the display never drifts, you'd be good to go.
 
#16 ·

Quote:
Originally Posted by mark haflich  /t/1533867/all-about-gamma#post_24788028


Would you be willing to pay for a transcript? Its easy to do yourself. Just play the audio into a voice recognizing typing program and then print the results out. you will need two lap tops or there are work arounds. You will probably need to edit a bit before the print out.
I would not, and I would most wouldn't. It's a question and suggestion. I'm more than happy to do without, but you can never be told yes to the question not asked.


Frank
 
#17 ·
Scott,


Thanks for arranging this interview with Joel. I must say, you are one energetic fellow! It seems every time I turn around you are producing more helpful and interesting material for us videophile/audiophile types. Perhaps we can get together again at CEDIA EXPO in Denver this September. I expect you will be roaming around the show floor interviewing more people. You have to stop and eat at some point, though.


One of the aspects of this interview I enjoyed the most was Joel's focus on human perception. I see often that discussions of gamma/EOTF in the forum fail to mention the connection with ambient light level and how humans perceive video images in varying lighting conditions. Video is based upon human visual perception, not the other way around. Many facets of how video should be implemented in an entertainment system are confusing for folks without a fundamental understanding of human factors. I first learned this from Joe Kane, Joel Silver, and their work in the Imaging Science Foundation back in the mid to late '90s. Anyone wanting to design a video reproduction system correctly must understand such fundamental principles. The ISF's original mission included a major focus on training their students how to design good systems, not just calibrate displays.


One thing Joel and I differ a bit on is the suitability of bias lighting behind larger displays. Even the International Telecommunications Union (ITU) recommends bias lighting usage with HD images occupying a 30 degree viewing angle, at only 100 nits screen brightness. The 30 degree viewing angle addresses optimum image size for 1920 x 1080 pictures. Actual display dimensions have no meaning without determining what the seating distance is. This is specified in ITU-R BT.710-4 'Subjective Assessment Methods For Image Quality In High-Definition Television'. As Joel pointed out, this is a recommendation, rather than a formal standard, which allows for differing opinions. There are differences between specific humans, of course, when it comes to their sensitivity to eye strain and viewing fatigue. However, standards bodies such as the ITU, SMPTE, EBU, etc., offer general recommendations that they have determined will apply to the majority of technology users.


There is another factor that needs to be considered in evaluating the impact of brighter, and high dynamic range (HDR), displays regarding viewing fatigue. Joel mentioned how the eyes can occasionally shift focus upon a dark perimeter around a 42" TV, while viewing a video program in a darkened room. He compared this to having a flashlight pointed into your eyes in the dark. This sudden, dramatic, shift in light level strains the eyes- causing us to typically flinch, squint, and/or blink. Our eyes are adaptive, but only up to a point. This sudden and dramatic change also occurs in actual video programs as well.


This characteristic of typical video program viewing can be demonstrated very easily. Any viewer can turn all the lights out at night when viewing a program, then turn to face away from the TV for a few minutes. It is common to see the room lit up by the screen, then strobe and throb as the light varies from scene changes. This is a technique used by program authors to keep the audience's attention. It stimulates the typical viewer's interest to alter the scene composition frequently through the use of varying camera angles, change of objects or people in the frame, varying light levels, etc. I have found that a greater factor in eye strain has to do with changes in image brightness more than image size.


Keep these interviews coming. It's so rare that average forum members get such exposure to world class authorities, experts, professionals, innovators, and practitioners.


Best regards and beautiful pictures,

G. Alan Brown, President

CinemaQuest, Inc.

A Lion AV Consultants affiliate


"Advancing the art and science of electronic imaging"
 
#18 ·
It's funny how the mind works. I edited my comments above to correct a misstatement. The "R" in "ITU-R, BT.710-4" above actually stands for the radiocommunications section of the ITU, not recommended practice. The "BT" stands for the broadcast television sub-section. This document is a recommended practice publication, but it's explained elsewhere to be such. Similarly, many are familiar with the ITU-R BT.709 document that contains recommendations for high definition broadcast. We see it popularly shortened to "rec709" in many discussions. Now I can go back to bed and actually go to sleep without this bugging me.
 
#19 ·
There's some historical issues in the interview. One was that gamma didn't become an issue until solid state (CCD) cameras became available. That's a couple decades off. While Image Orthicons (IO) and vidicons had inherent logarithmic correction, the Plumbicon (and later Saticon) tubes that became available in the early 60s had a unity transfer characteristic and therefore required electronic gamma correction. The analog method of correction utilizes the logarithmic relationship between the current and the voltage on a semiconductor junction. A common approach to vary the gamma was to over correct it and then use a mixer (in early designs just a pot) to vary between linear and log corrected video. Setting the amount commonly utilized a "chip chart" which contains neutral color exponential steps. As log correction would create infinite gain at black, as well as practical limitations of the gamma correction circuit, there was a tapering to linear in the lower levels. It wasn't until digital processing was needed, especially as 601 became common, that emulation of these analog techniques in digital were required. This required a more precise definition of where the linear section began, which with 709 is around 8% above black post gamma correction.


The exponential signal-to-light function of a CRT is from the triode transfer characteristics.


One thing I wonder about the Dolby chart's measure of the visibility of quantizing errors is if it included dithering. Also what they are pointing out, which isn't visible in the video, is that these errors are mostly in the darkest regions. Visual perceptual sensitivity has greater non-linearity than traditional gamma. Also, in order for 8 bits to be properly utilized, the precision of the display processing needs to be higher with more bits.


A pure color is referred to as monochromatic. Green is the greatest challenge as it's in the middle of the visible spectrum while the other primaries are at the edges. Purer primaries have less crosstalk into the adjacent cone spectral sensitivities - Green to L & S, Red to M & S and Blue to M & L.


As Charles Poynton has pointed out, we were indeed fortunate that we followed the path of compensating for the non-linearity of the CRT at the source with gamma. The gamma curve was close to visual intensity perception and allowed quantization to 8 bits to be practical for video. Assuming the maximum gain in the linear section is around 4.5, an extra 3 bits would have be required if linear had been adopted, with much of the extra information in the highlights wasted.


While a better locked down system compatible with existing standards is a significant improvement, better correlation to visual perception would be a step forward. Images which fill peripheral vision with resolution to the limits of perception, colors to the limit of the locus (Rec 2020 is likely good enough), high dynamic range and frame rates that vision limits for motion perception is the goal. Maybe next year?
 
#20 ·

Quote:
Originally Posted by TVOD  /t/1533867/all-about-gamma#post_24793809



The exponential signal-to-light function of a CRT is from the triode transfer characteristics.

Exactly - it's a common misconception that the "gamma" of a CRT comes from the phosphor response, but the phosphor response to beam intensity is actually quite linear, at least for the first 80% of the rise function in luminance. As you point out, the non linearity comes from the relationship between applied voltage and beam intensity.

Quote:
Originally Posted by TVOD  /t/1533867/all-about-gamma#post_24793809


As Charles Poynton has pointed out, we were indeed fortunate that we followed the path of compensating for the non-linearity of the CRT at the source with gamma. The gamma curve was close to visual intensity perception and allowed quantization to 8 bits to be practical for video. Assuming the maximum gain in the linear section is around 4.5, an extra 3 bits would have be required if linear had been adopted, with much of the extra information in the highlights wasted.

Yes, I was a bit surprised at Joel Silver's characterization of the CRT's nonlinearity - he made it out to be some sort of defect, when in fact it was a very "fortunate coincidence" that the voltage to luminance relationship of a CRT was very nearly the exact inverse of luminance lightness relationship of the human visual system. As Alan just pointed out in this thread, human perception is paramount. The fact that with a CRT you'd have a perceptually uniform "ladder" as you increase the video input level means that you can actually avoid quantization artifacts, such as banding, with a mere 8 bits!


Linear light is great, once you're at about 16 bits or so.


The concept of perceptual uniformity also helps to understand how BT.1886 works. As far as I can understand, BT.1886 accomplishes two things:


1) Standardizes the exponent of the EOTF (2.4)


2) Changes the shape of this function depending on the black level of the display. This is important, as our visual systems become less sensitive to changes in light when these changes occur at higher luminances. This means that if our display has a relatively high black level, we are going to be less sensitive to changes in luminance that occur at the low ends of our video signal. BT.1886 takes our black level into account, and accelerates the growth of the luminance function in the lower ranges to compensate for the raised black level. Without this compensation, we'd get crushing of our "blacks".


Here's a thread on BT.1886 and these issues.


One more quibble in an otherwise fantastic episode:


In the figure shown at 7:20, the Y axis should read luminance, not brightness. Brightness is the perceptual correlate of luminance, and if the figure actually represented brightness, the blue CRT curve would be a straight line.
 
#21 ·

Quote:
Originally Posted by Stereodude  /t/1533867/all-about-gamma#post_24779165


I also thought it was strange that Joel presented the information that BT.1886 is not suitable for a bright room, only for a dark room. BT.1886 can compensate for an elevated non-zero black level. If you measured and calibrated to BT.1886 in a bright room, wouldn't BT.1886 be still correct? I do realize the bright room BT.1886 curve will not look like the BT.1886 curve if you measured and calibrated in a dark room.

BT.1886 takes the black level of the display into account, not the ambient light conditions. Joel is correct - with a brighter room, adaptation means that the image will appear darker. In order to compensate for this, you need to lower the exponent of your function. Now there's no reason that you can't use a BT.1886-like function that uses a lower exponent, such as 2.2, or in extreme cases, 1.8!
 
#22 ·

Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24796257


This means that if our display has a relatively high black level, we are going to be less sensitive to changes in luminance that occur at the low ends of our video signal. BT.1886 takes our black level into account, and accelerates the growth of the luminance function in the lower ranges to compensate for the raised black level. Without this compensation, we'd get crushing of our "blacks".
If a display is calibrated properly for its best black level, and then subsequently the black level is raised, doesn't this also decrease the amount of gamma correction (assuming the black level is raised ahead of the gamma correction)? Thanks for the link to the other thread, got some reading to do.
 
#23 ·

Quote:
Originally Posted by TVOD  /t/1533867/all-about-gamma#post_24796471


If a display is calibrated properly for its best black level, and then subsequently the black level is raised, doesn't this also decrease the amount of gamma correction (assuming the black level is raised ahead of the gamma correction)?

If I understand your question, I think it depends on what exactly is going on when the black level drifts up, or is raised. If the black level is raised uniformly across the entire signal range, then the curve just shifts upwards, and the shape of the function remains the same.


If the lower range shifts up more than the mid-upper ranges, then I think that this will accomplish something similar to what BT.1886 does.


The problem is that when studios and consumers calibrate their displays, they need a particular luminance function to target. Suppose the target is a gamma of 2.4 - typically what is done with a raised black level is to simply shift the function up: Luminance = (Input^2.4)+black offset - this is often referred to as an "output offset".


The BT.1886 function shifts the offset to the base part of the equation: Luminance = (Input+offset)^2.4. - this is referred to as an "input offset".


Btw, the relevant equations can be found here , in Annex 1.


So, yes, if we calibrated our displays with a perfect black level to an exponent of 2.4, and our displays naturally drifted in a way that preserved perceptual uniformity, and we just left things as is, we might be ok. But we need a way to actively characterize the EOTF as a function of black level, so that we can generate appropriate LUTS/calibration targets.
 
#24 ·

Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24796537


If I understand your question, I think it depends on what exactly is going on when the black level drifts up, or is raised. If the black level is raised uniformly across the entire signal range, then the curve just shifts upwards, and the shape of the function remains the same.
I was thinking of a technique used to vary gamma, but then I remembered a part of that was to also vary the gain ahead of the exponential correction.
Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24796537


The BT.1886 function shifts the offset to the base part of the equation: Luminance = (Input+offset)^2.4. - this is referred to as an "input offset"..
As the offset increases, the values in the highlights would increase faster with the exponential correction compared to the lowlights. Depending on the processing headroom, that could cause clipping.


What if display manufactures implemented the equation (((1-offset)*Input)+offset)^2.4, essentially keeping the Input value of 1, a normalled point, constant. For instance, if the offset was .1, then (((1-.1)*1)+.1)^2.4=1. I think this would also have the effect of lowering the gamma correction with added positive offset, but I don't know how well it would track visual sensitivity.
 
#25 ·

Quote:
Originally Posted by TVOD  /t/1533867/all-about-gamma#post_24810161


As the offset increases, the values in the highlights would increase faster with the exponential correction compared to the lowlights. Depending on the processing headroom, that could cause clipping.

Not sure if you've got this right:


This is what the BT.1886 curve looks like as a function of varying the black level. Here are a family of curves with black levels ranging from 0% to 10% of max luminance. Note that when the black level is a perfect black (0% luminance), it's a perfect 2.4 gamma curve.





Also note that the offset in the BT.1886 isn't a linear function of the actual black level. Here is a plot showing the relationship between the measured black level of a display, and the corresponding "offset" value used in the BT.1886 function:





In contrast, what was typically done before BT.1886 was a ranged and scaled approach, where the desired gamma function was fit between the min and max luminance of the display. See below:




Note that, unlike in the BT.1886 function, the shape of the function retains its characteristic curve across different black levels. With a naive understanding of the issue, this may seem desirable, but it fails to take into account the perceptual issues discussed earlier. As a result, the image is crushed.

side note: In fact, if you look carefully, with higher black levels, the curve rises even more slowly at the beginning compared to with lower black levels, so you get even more crushing! (you get crushing of luminance and crushing of brightness (the perceptual correlate of luminance)

Quote:
Originally Posted by TVOD  /t/1533867/all-about-gamma#post_24810161


What if display manufactures implemented the equation (((1-offset)*Input)+offset)^2.4, essentially keeping the Input value of 1, a normalled point, constant. For instance, if the offset was .1, then (((1-.1)*1)+.1)^2.4=1. I think this would also have the effect of lowering the gamma correction with added positive offset, but I don't know how well it would track visual sensitivity.

This is what your idea looks like as a function of black level:




Your equation produces severe clipping at the low end. When the black level is 10% of peak luminance, the function produces a min luminance of 0.4% of peak luminance.
 
#26 ·

Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24810552


This is what the BT.1886 curve looks like as a function of varying the black level. Here are a family of curves with black levels ranging from 0% to 10% of max luminance. Note that when the black level is a perfect black (0% luminance), it's a perfect 2.4 gamma curve.



Also note that the offset in the BT.1886 isn't a linear function of the actual black level. Here is a plot showing the relationship between the measured black level of a display, and the corresponding "offset" value used in the BT.1886 function:

I think this is due to the exponential gamma correction. On my equation to make constant output steps would be:

(((1-(offset^(1/2.4)))*Input)+(offset^(1/2.4)))^2.4
Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24810552


In contrast, what was typically done before BT.1886 was a ranged and scaled approach, where the desired gamma function was fit between the min and max luminance of the display. See below:



Note that, unlike in the BT.1886 function, the shape of the function retains its characteristic curve across different black levels. With a naive understanding of the issue, this may seem desirable, but it fails to take into account the perceptual issues discussed earlier. As a result, the image is crushed.
This looks to be changing gain and adding an offset after the exponential gamma correction. My proposal is basically the same thing but before the gamma correction instead.
Quote:
Originally Posted by spacediver  /t/1533867/all-about-gamma#post_24810552


This is what your idea looks like as a function of black level:



Your equation produces severe clipping at the low end. When the black level is 10% of peak luminance, the function produces a min luminance of 0.4% of peak luminance.
Not sure how that occurred. When the video is 0, the offset provides a Y value. Maybe incorrect grouping? Anyway I ran the curves with my above equation with 0 to .1 with .01 steps:




Look close?
 
This is an older thread, you may not receive a response, and could be reviving an old thread. Please consider creating a new thread.
Top