Home of Professional Cinematography since 1996

style="margin-bottom: 0">
Color Bit Depth

Published : 31st August 2004

The point was made and the question asked that....


>what is the best way to get around Mach Banding when going from >16bit 2k 4:4:4 images and downrezzing them to a 8bit HD(1920x1080 >24PsF) 4:2:2 D5 for Digital Cinema?

There is no real substitute for color bit depth. There are bandaids. Dithering, dynamic rounding, etc. are bandaids. 8 bit color is a relic of television, and is an outgrowth of the capabilities of phosphorous based display technology. Since phosphorous monitors are inexorably becoming archaic, we should consider obsoleting the technology of 8 bit color along with it. Who was it that said that Digital Cinema must abandon all links to Television? I think this is part of what he meant.

Once again, let me urge; if we are going to re-invent the future, we ought to remember to make it better.

Dave Stump ASC
VFX Supervisor/DP
LA, Calif.



David Stump wrote :

>8 bit color is a relic of television, and is an outgrowth of the capabilities >of phosphorous based display technology.

Yes, but most current LCD panels are 8 bit... at least the ones now shipping.

Jeff Kreines



>Yes, but most current LCD panels are 8 bit...at least the ones now >shipping.

Precisely my point. Now is the exact moment to "cut the cord".

Dave Stump ASC
VFX Supervisor/DP
LA, Calif.



Dave Stump writes :

>Since phosphorous monitors are inexorably becoming archaic, we >should consider obsoleting the technology of 8 bit color along with it.

Without disagreeing, on a practical level we find no end to the need to make images "8-bit proof"; to make sure that if they're taken through an 8-bit step, they won't look terrible. Otherwise one will be more often than not disappointed with how they're presented. You want it to look good on ATSC, DVD, etc.

Tim Sassoon
Sassoon Film Design



Tim Sassoon wrote :

>Without disagreeing, on a practical level we find no end to the need to >make images "8-bit proof";

Yes, very true. And it will remain true for a long long time.

>to make sure that if they're taken through an 8-bit step

Ah, now that's a good one! Step indeed!

Jeff Kreines



> Precisely my point. Now is the exact moment to "cut the cord".

---Precisely....agreed!
"All ties to television as we now know it.....must be irrevocably CUT."

Martin Euridjian's statement to that effect needs to chiselled in stone somewhere in a very visible place in every city where production and post production play a major role in the economy.

Jeffery Haas
Freelance editor, camera operator
Dallas, Texas



>> most current LCD panels are 8 bit

> Now is the exact moment to "cut the cord".


At the same time, 8 bits per channel is well below the human threshold of color-difference perception. In other words, the average observer cannot distinguish between a significant portion of the colors that one is able to represent and display with a well tuned 8-bit-per-channel RGB system.

The first generation 23 inch Apple HD Cinema Display has a delta-E (measurement of perceivable color difference) of about 0.5DE. A good CRT can't do much better than 1.0DE. The untrained observer can see 1.3DE and above. A trained observer will see below that, say, about 1.0DE.

In a nutshell, 8 bits per channel on a modern high-quality LCD -- properly driven -- can produce better images than the average human visual system can "see". Here the word "see" is in quotes because color vision is a psycho visual affair, not a matter with many absolutes. What you "see" is as much a function of what your brain decides to let you see as what the eyes actually capture.

Having said that, 12 bit per channel LCD's look amazing.

Martin Euredjian
eCinema Systems, Inc.
www.ecinemasys.com



Martin Euredjian wrote :

>...In other words, the average observer cannot distinguish between a >significant portion of the colors that one is able to represent and display >with a well tuned 8-bit-per-channel RGB system.

I heard recently at Cinegear CML Saigon Annies (where all truths are expounded over salt broiled shrimp and beer which in turn makes said statements seem more true) that studies show the average threshold equals around 11 to 12 bits per channel. That there's a small segment of the population that can differentiate colors at the 14 bit level, but most of the bell curve rolls off just under 12.

I bet the 8 bit is fine under less than ideal viewing conditions. Maybe on seeing the test patterns on a large, decent display and in a darkened room, the viewers can actually distinguish them better ? I just cannot imagine that 8 bit is at our ideal threshold.

I admit I only absorbed perhaps 20% of the Saigon Annie discussion on color space, but even that much was fascinating (not to mention the politics involved!).

Our industry does need to come to some consensus with regards to these issues so we can adopt large, future proof color space and more interchangeable file formats for post. It would certainly help DP's going into projects where we never have time to test thoroughly the workflow that in the end provides us with its own version of the colors to be projected theatrically. Heck, often we can't get an answer on where the cameras are coming from and where the filmout will be until 2 weeks out and tech scouting !

Funniest quote at Annie's: "The studios say, 'if 12 bits per channel looks great, then 16 bits must look better. We want all our stuff in 16 bit!' But for what ? They want to pay 8 bit prices."

Mark Doering-Powell
LA based DP



>At the same time, 8 bits per channel is well below the human threshold >of color-difference perception

While this may be a commonly held belief, it does not hold up to rigorous testing.

In October 03, we (under request from DCI) did a comprehensive discrimination threshold test using theatrically projected images - at 40, 4, and 0.4 cd/m^2 using a gamma of 2.6. These were projected images on a 45' screen at the DCL in Hollywood. We tested approximately 75 subjects, from both "expert viewers" and non experts. The test patterns created steps in luminance corresponding to 1 count in 8,9,10,11 and 12 bits. The viewer was asked to identify the orientation of the test pattern, to verify if he really saw the steps.

The results of the experiment will be published in the September 04 SMPTE journal, but can be summarized by the following :

* Gamma 2.6 best follows the threshold of human perceptual discrimination
for luminance at cinema levels
*  All viewers could see 8 bits all the time
*  Almost all observers could see steps at 10 bits
*  Some observers (40%) could see steps at 11 bits
*  Essentially no observers (1%) could see steps at 12 bits.
* Observers learned to see steps and became more sensitive with repeated viewing

These differences calculated to very small DeltaE* - in the smallest case, the threshold of discrimination was 0.05 DE* (if I remember correctly), much smaller than previously thought to be visible.

This was the experiment that caused DCI to adopt a 12 bit (gamma 2.6) recommendation.

If the display does not exhibit contouring at 8 bits, this will be caused by several factors :

* Noise - (either in source or display) dithering will hide contours very effectively

* Environment - environmental contamination of the display (e.g. lights on around display)

Note that if you translate this into linear space, it requires a lot more bits.

Matt Cowan



Why doesn't digital cinema use 10 or 12 bit 4:4:4 such as the Sony HDCAM SR as it's master rather than the inferior D5 source? It is the Sony format just too new?

Jamie Tosi
Digital Media Specialist
Santa Monica, CA



Matt Cowan wrote:

>Observers learned to see steps and became more sensitive with >repeated viewing

And once viewers learn to see artefacts (like digital compression or noise reduction artefacts) they become pickier, and the bar is raised.

Up to the limits of the eyes and brain, that is...

Jeff "2-bits ain't worth 2-bits anymore" Kreines



Matt Cowan wrote :

>Note that if you translate this into linear space, it requires a lot more >bits.

So that 12-bits was actually 12-bit log or 12-bit linear?

Just curious to know if 10-bit Log DPX files are going to be enough for the digital projection standards.

Jason Rodriguez
Post Production Artist
Virginia Beach, VA



>>At the same time, 8 bits per channel is well below the human >>threshold of color-difference perception.

>While this may be a commonly held belief, it does not hold up to >rigorous.


Rigorous? How's 10,000 test subjects with the testing being conducted by the Munsell Color Laboratory in Rochester NY by the Color Scientists that literally wrote all the books?

I don't have the time to do it right now. I'll post a few images later.

>In October 03, we (under request from DCI) did a comprehensive >discrimination threshold test using theatrically projected images - at 40, >4, and 0.4 cd/m^2 using a gamma of 2.6.

I'm not sure what this tested. What does projector gamma have to do with color perception?

>steps in luminance corresponding to 1 count in 8,9,10,11 and 12 bits.

Luminance!? So, you were just going up and down the L axis in CIELAB? Of course you'll get better delta-E's!

Martin Euredjian
eCinema Systems, Inc.



Probably important to go back to part of what I said. I find that people can read through a post quickly and miss any subtleties or precision in the language :

>the average observer cannot distinguish between a significant portion >of the colors that one is able to represent and display with a well tuned >8-bit-per-channel RGB system.

This does NOT mean that ALL color differences representable in 8 bit RGB are tough to see. That is NOT what I said, right?

We've all seen good examples of 8 bits gone bad. I remember an old Alpha Image video switcher that had the ability to change the output processing from 8 bit truncate to 8 bit rounded. Huge difference for some passages, none for others. With truncation some finely shaded backgrounds had all sorts of very visible banding. It just so happened that the color differences were large enough that almost anyone could pick them off.

Personally, I think that 8 bits per channel --with the right processing -- is excellent for viewing by the masses.

>8 bits per channel on a modern high-quality LCD -- properly driven-- >can produce better images than the average human visual system can >"see".

Again, qualifiers: "high-quality LCD", "properly driven", "can produce", "average"

>color vision is a psycho visual affair, not a matter with many absolutes.


No need to explore that further right?

Martin Euredjian
eCinema Systems, Inc.



Matt Cowan wrote :

>So that 12-bits was actually 12-bit log or 12-bit linear?


No it was 12 bit gamma *corrected* linear of '2.6' which is neither, and I've no doubt you'll need to know the range of luminance this was tested over, the conditions of viewing (average screen luminance, surround, environment, distance from screen, size of pattern, size of feature within the pattern, how noise free was the system, how were the images projected on screen, was the screen perforated or not (a killer of sharpness) etc etc.) Was it a colour based assessment or just a tone scale.

>Just curious to know if 10-bit Log DPX files are going to be enough for >the digital projection standards.

In the absence of noise, 10-bit Log DPX files are not enough, according to conversations I've had with somebody involved, based upon the tests carried out.

FWIW, I do find it a little strange that some of the proposed systems and methods and the corresponding tests are always hard to get the details on. They are all carried out as you would like to be, but obviously not everybody can be there at the same time. If there was truly freely available scientifically based (i.e. peer reviewed, mostly bias free) set of procedures etc, I think there would be less of an issue with some of the standards proposed as you could if so inclined go and do the same test yourselves

(Note : I'm probably in a position to know some of the people involved so I'd like to point out its not a slight on individuals more the crazy way our industry works)

As somebody mentioned earlier, studios want 16 bit for 8 bit prices, well I'd say they also want 4K images for less than 2K prices, it is the economics of today. So what is really important is to provide a migration path to this 'ideal' world of the "new standard". its quite funny the number of clients who want 4K, must have 4K, etc but then you say, have you got 3-4 times the money? or 3-4 times the schedule.

These are of course things that change overtime.

| Kevin Wheatley, Cinesite (Europe) Ltd | Nobody thinks this
| Senior Technology | My employer for certain |
| And Network Systems Architect | Not even myself |



>Kevin Wheatley asked for details of the discrimination threshold test.

Here are the design parameters of the test. I suggest that anyone interested read the entire paper in the SMPTE journal when it was published in September. The delay in publication relates to lead times for print. This work was presented at the SMPTE conference in November (30 days after the test was complete) and presented again in January at a DC-28 technology meeting in January 04.

1/. The experiment was designed and performed by myself, Glenn Kennel (formerly of Kodak now with TI - and principal developer of the Cineon format). Dr. Tom Maier, Kodak, Brad Walker, TI, and under the guidance of the SMPTE DC 28 color ad hoc group.

2/. The intent of the experiment was to determine what the human is capable of seeing, in a theatrical context. That means dark surround, large screen images with a range of viewing distances relating to "normal" theatres. This drove the choices for experimental design

3/. The basic experimental parameters were calculated from the threshold model developed by Barten (P. G. J. Barten, Contrast Sensitivity of the Human Eye and Its Effects on Image Quality, SPIE-The International Society for Optical Engineering: Bellingham, WA, 1999.)

4/. This test was for tone scale (luminance) not color.

5/. The test measured luminance modulation (in cd/m^2) as opposed to bitdepth. Bitdepth was calculated afterward, and could be related to any desired coding scheme. Gamma 2.6 was chosen because it required the minimum number of bits of any of the "uniform" schemes.

6/. The test pattern was a low modulation square wave of 16 pixel duration, with 13 repetitions in a test square of 208 pixels square on the screen. Pixel size was 0.213" square on the screen

7/. Subjects were grouped at viewing distances of 27, 41, 57, and 77' from the screen. (you can calculate the spatial frequency of the pattern....)

8/. Luminance’s tested were approximately 75%, 7.5% and 7% of cinema brightness

9/. Projection was from an internal pattern loaded into a DLP Cinema projector, providing 16 bit quantization in pattern luminance. The noise in this environment is quite low.

10/. Surround of the pattern was 90% of the luminance of the pattern itself, filling the theatre screen. Theatre surround was "normal theatrical"- dark.

11/. When we calculated the log coding equivalent, we needed an extra bit to avoid visibly quantised thresholds. (it appears at the bright end of the tone scale)

All this said, relatively small amounts of noise provide masking of contour artefacts. Current digital cinema releases are coded in 8 bit MPEG (at 65 Mbits) and most of the time, contouring isn't visible. (but sometimes, it is.)

Matt Cowan.



Martin,

Please give me a one or two line explanation of ODE. I realize it is probably vastly more complex than one or two lines of exposition, but I believe I have a basic grasp (that would definitely benefit from some ODE 101).

Brent Reynolds
Producer / DP
August Moon Productions
Tampa, FL

P.S. - Of course, anyone else feel free to answer - it's just that Martin brought it up - so I thought I would ask him.



Martin Snashall sent this to me, which I thought might be of interest. (He's referring to the Abekas A64 and A84, both of which he designed.)

Well, things come back round to haunt us again. The "rounding" and "truncation" Martin saw on the Alpha image switcher was originally put on the A64, consequently the A84, and Alpha image pinched it from there!

Quantel had this thing they called "dynamic rounding", which despite all their protestations about it being something they had invented, was just dither. The basics of dither is that is you multiply an 8 bit signal by an 8 bit control, then you get a 16 bit result, which then needs to be taken back down to 8 bits for display. What dither does it to take the 8 bits you would throw away, and add a random number to those 8 bits. If the result produces an overflow, then you add one to the 8 bit display.

What this does is to produce an output that over time is closer to the actual result, rather than just the truncated result. This technique is another one that seems to have been lost in this computer age, as a fixed input frame, when, say, faded to 72%, would not produce a fixed output frame, as nearly all the pixels will be dithering by 1 bit. In fact, adding noise usually produces a better result visually.

In many cases this is more important and visible in a YUV system, especially on blue colors. The reason is that the bits have very uneven weight in YUV, as shown by the blue vector. If you have a fairly saturated blue, then a single bit change in the luminance can alter then output blue level by some 4% (this is due to blue only contributing 11% to the luminance signal), assuming that the U & V stay the same.

Jeff Kreines
Kinetta



Jeff Kreines writes :

>In fact, adding noise usually produces a better result visually.

Which unfortunately reduces compression efficiency. And keep in mind that one only needs to apply .5% noise to make up the difference between 8 and 16bpc - hardly visible, but can make a 30% larger file. There may be some value in adding noise at the projector - unmentioned in DCI, though the idea's been around for a while.

Tim Sassoon
Sassoon Film Design



> Please give me a one or two line explanation of ODE.

Here's a simple one. I may be able to elaborate later...I have back-to-back meetings and very little time.

Take a 50% gray card, continuous tone. Now, print the same 50% gray as a newspaper would (by using black dots on white paper). You'd have 50% of the surface area covered with black dots and the other 50% would be white.

The difference from 50% gray to either black or white is 50 delta-E.

This also opens the door to another subject. As you know, if you are sufficiently far away from both of the above cards you'll see exactly the same 50% gray.

The CIELAB color space has a lot of advantages in terms of uniformly describing color differences. Imagine a centre trunk that represents the "L" portion, the brightness component. As you move up and down you move from black to white. At any level you may choose, say L=50, moving horizontally away from centre produces greater color saturation. One can draw small ellipses at any horizontal level to enclose all colors that fall below the threshold of human color-difference perception.

Anything within the ellipse looks the same, even though a laboratory instrument would prove otherwise. Differences greater than the ellipse can be seen by the average observer. As you get farther and farther away from centre (more saturated colors) the threshold increases (the ellipse gets larger). This means that the average delta-E is smaller as you get closer to less saturated colors and greater for more saturated colors. Or, if you will, one can see smaller changes with black and white than with saturated colors.

Here's another topic to research and become familiar with : Metamerism. Do a Google/Yahoo search to open-up a world of confusion.

Martin Euredjian
eCinema Systems, Inc.



FYI, it is generally (but not universally) accepted that a human can differentiate a 1 percent difference in luminance. Hence the problem that, in a 12 bit linear system for example, you have a 1 percent difference between step 100 and 101 (which is approximately a 41:1 ratio with step 4095) but a 4 percent difference between step 25 and 26. The former is generally not visible but the latter most certainly is.

That is why the low lights tend to be problematic and why we use non-linear systems where possible.

Tom Tcimidis



> reduces compression efficiency.
> .5% noise can make a 30% larger file


I knew the first part (and I always talk about it) but never bothered to quantify it in some way. Have you run any test over a wide range of frames?

Martin Euredjian
eCinema Systems, Inc.



Thanks Martin,

I was thinking along the right lines, but your explanation was a big help.

Metamerism - huh?

Whenever get a little too big for my bitches and think I actually I know something. I just come here and ask a question.

I'm not really a DP, I'm a monk and this forum reminds me of my vow of humility.

Brent Reynolds
DP / Monk
Tampa FL



And in fact dithering is SOP in digital (audio) mixing, and for the same reasons.

The (endless) debate there is over not if to do it, but how to do it.

Sam Wells



> 4. This test was for tone scale (luminance) not color.

Makes sense. I was talking color. Of course, the threshold for greyscale discrimination is lower. I imagine it's part of the "survival in the wild" business.

Martin Euredjian
eCinema Systems, Inc.



Martin Euredjian writes:

>I knew the first part (and I always talk about it) but never bothered to >quantify it in some way. Have you run any test over a wide range of >frames?

Well, to be more precise, wouldn't it be .39% noise (100/256) to cover the 8-9 bit gap? Which (and it must be calculated in 16-bit space) covers the rest as well. In practice, .5% noise (uniform or gaussian) tends to look more like the original 10+bit, which may be more a software artifact than anything else. I just did a test to quantify it, first using Marcie in log, then with a horizontal ramp, both to .jp2 at 90% quality out of Photoshop CS with the Orphanage log ICC profile. What I found, unsurprisingly, is that the amount of difference changes with the compression ratio. At 90% it was about 10%, at 80% it went up to a 25% differential, etc. I was probably being excessively emphatic before, but there is a real difference. Marcie has quite a bit of flat color in the LAD patch, etc.

Dave Stumps writes :

>Also, grossly clipped areas can solarize, and even invert in color. Very >difficult to fix in post

Not to mention chroma fringing on the highlight. IMHO digital origination cannot replace film until there are cameras with similar highlight response. It should be our primary issue.

Tim Sassoon
Sassoon Film Design



>...the Orphanage log ICC profile...

Which, thanks for reminding me, has a home now:

http://www.theorphanage.com/tech/

Stu Maschwitz,
CTO, The O, SF/LA



>Whenever get a little too big for my bitches and think I actually I know >something. I just come here and ask a question.
>I'm not really a DP, I'm a monk and this forum reminds me of my vow of >humility

---Amen...this list is better than a four year curriculum at USC Film School.

Jeffery Haas
freelance editor, camera operator
Dallas, TX



Kevin Wheatley wrote :

>In the absence of noise, 10-bit Log DPX files are not enough, according >to conversations I've had with somebody involved, based upon the tests >carried out.

Most digital cameras as well as film have a certain amount of noise though, so this should be fine.

I mean right now aren't most DI's being done with a 10-bit log scale, and even HDCAM-SR with the Genesis will only be 10-bit log.

I'm assuming that banding will only be a problem when digital projection is in full swing, since film projection and film-outs should mask any banding artefacts, correct?

Jason Rodriguez
Post Production Artist
Virginia Beach, VA



>Metamerism - huh?

Two objects may appear to be (and/or reproduce) the same colour under one light source but different under a different light source.

It's to do with the fact that all visual and recording systems encode a continuous spectrum in terms of the proportion of three primary colours. But those colours aren't always the same: they may each occupy a narrow or broad part of the spectrum depending on the system, and may therefore be affected by different profiles of light source.

Dominic Case
Atlab Australia



>That is why the low lights tend to be problematic and why we use non->linear systems where possible.

And one of the reasons for perceptually-uniform color spaces, like CIELAB (uniform-ish), being more useful.

Martin Euredjian
eCinema Systems, Inc.



Dominic Cased writes:

>Metamerism - huh?

Two objects may appear to be (and/or reproduce) the same colour under one light source but different under a different light source.

The brownish sheen that deep blacks from some ink-jet printers unintentionally make is an example. It brings up the point that we're thankfully working both for and in a transmissive medium digitally. IMHO prepress color can be much more difficult. It's pretty hard to simulate a foil stamp or a gloss varnish on-screen without a lighting context.

Tim Sassoon
Sassoon Film Design



Jeff Kreines writes:

>Martin Snashall sent this to me,....

>.... The basics of dither is that is you multiply an 8 bit signal by an 8 bit >control, then you get a 16 bit result, which then needs to be taken back >down to 8 bits for display.

Dithering is no news to audio folks. It's always been the preferred method of bit-depth conversion and level adjustment, and results in more natural-sounding digital audio, with less audible granularity. With that said, some dithering algorithms are better than others.

Most of them these days are pretty good, but it wasn't always that way.

Dan Drasin
Producer/DP
Marin County, CA



Jason Rodriguez wrote :

>I mean right now aren't most DI's being done with a 10-bit log scale, >and even HDCAM-SR with the Genesis will only be 10-bit log.

As far as "most DI's" I wouldn't know, but from the standpoint of 10bit log files produced from a film scanner today, your probably talking about a CCD A/D of '14 bits linear' which is sampling the film via passing light through it, which is not a linear capture device with regards to scene exposure due to the non-linear response of the film to intensity. (After that you generally 'log' the data in the electronics, subtract sensor noise, do some temperature based cancellation and other things like a matrix to compensate for the imperfect nature of filters and so on, to get a picture)

The digital camera approach is more straightforward and gives you a truer number of bits of linear, i.e. there is no optical compression of the tone scale before the CCDs, but you may have only 12 bits depending on your sensor. you then perform similar electronic calculations to get your self a picture.

What this means is that for a given amount of dynamic range and A/D bit depth, and the same amount of noise and high enough level of electronic bit depth for processing, I'd expect to see more bands in a straight CCD capture system.

This is not to say that the film compression is ideal (perceptually perfectly uniform), or that the pictures can't look great out of either (or terrible as hell for that matter)

>I'm assuming that banding will only be a problem when digital >projection is in full swing, since film projection and film-outs should >mask any banding artefacts, correct?

Yes, although the grain from 5242 (and laser recorders in general) and 2383 print won't mask everything, we have, in the last 11 or so years, had 3 or 4 problem sequences for VFX work where banding was visible even though minimal grading was done (basically scan and then record). Some of these occurred on older intermediate and print stocks than are currently used, these would have had "more grain" than today. I would say that it is probably a very small percent where it becomes a problem, rather than a 'look'.

Moving to digital projection will certainly make the percentage of problems shots increase for this reason amongst many

| Kevin Wheatley, Cinesite (Europe) Ltd | Nobody thinks this |
| Senior Technology | My employer for certain |
| And Network Systems Architect | Not even myself |