Some people are still being told that 200:1 is all we can see.... - Page 9

But you're still confusing the two things together. The compression will occur regardless of whether gamma is even implemented in the system. Do you see my point? Take gamma completely out of the system: pretend everything is linear (as in a direct film capture). There is still compression/alteration of the tone scale going on. This is distinct from gamma, which is my point. And also what Poynton is basically explaining. That's why he isn't discussing gamma at all in the relevant section. You are inserting gamma as being part of that section when in reality it is completely absent, and that's why you're fixating on gamma as being relevant in this when it's basically not relevant.

This is exactly what I was describing with regards to film's S-shaped gamma.

It's why I included the modifier "per se" in my statement that gamma is not being implemented in film, and also why I called it a "gamma-correction flow" rather than just "gamma." I.e. you are not imposing a non-linearity to the signal and then inverting that later on as in the gamma-correction flow in video which is what we are discussing.

Please, at least try reading what I write. Otherwise you're just continuing to misrepresent me as you have always done.

No I don't see the same, but again as I said before this test is poor because it is so display-dependant. It doesn't really prove anything in terms of your visual acuity because the limitation in most situations will be the display, not your vision. If you view this on a low-ANSI display (like a CRT), I predict it's unlikely you'll be able to see the 4% square, perhaps at all even if you're looking at it.

The test Darin and I suggested with a checkerboard and shadow puppets is more relevant, though it too is display dependant in terms of disproving the claims. If you have a very low ANSI display that maybe only gets 80:1 and you can see puppets, it only really proves that we can see more than 80:1 which is below the claims being made here. On the other hand if you have a display system with an ANSI of say 500:1 and you can see the puppets, then that indicates you can see more than 500:1 across a scene, and that disproves some of the numbers being suggested in this thread which are below this value.

Chris: The 5-square test demonstrates empirically that a 3-4% bright square simply vanishes when you superimpose a white square on now two LCD monitors. That square is projected onto the screen, but two humans see it disappear. Maybe more people should do the test (but you need to adjust the square brightness to suit your display, I can resolve a 2% brightness square).

P.S.: Are we getting a reference for 1000:1 or not? Or are you now down to >500? Come on, be constructive and help us out here with more than conjecture. We want a reference for your claims. Just bashing everybody else doesn't automatically make you right.

But the luminances of these squares and the amount of spill depends on the display system. Spill in the system will obscure the visibility of these blocks. That's why it's not particularly relevant, because some displays may cause the squares to be more or less visible. It does not isolate anything to the human vision.

Something a bit more interesting than the 5 square test that I beleive makes the same point.

http://www.michaelbach.de/ot/lum_dyn...ast/index.html

Chris: Let me try and understand. How will spill in the sytem (whatever that means) obscure the visibility of the squares? Those squares are visible simply by moving the fovea away from the bright square. Then you say it's not particularly relevant. Is it not relevant at all? What do you mean by "particularly" relevant? I agree that the impact will be different across displays, because everybody has their particular contrast and brightness settings. But everybody's LCD screen is likely below 1000:1 CR. So if our human vision can discriminate this CR, then is absolutely no reason why we shouldn't be able to see a 3-4% bright square. It shouldn't just vanish and reappear when we move about the image. You really ought to be able to explain this phenomenon. My interpretation is that we are moving up the CR elevator on the full luminance scale and faint objects simply vanish when we do that.

mdtberi: Great link!!!! Cool site, it's going into my permanent bookmark list.

Hey. How's it going? I'm Matt Trentacoste. I keep seeing my thesis cited and figured I might as well jump in and join the fray.

Thanks for the link! A really interesting collection of visual illusions... thought provoking regarding the concept of simultaneous contrast and object/motion/color perception.

Hi Matt,

Welcome to the forum. I'm interested in your take on an experiment like the one I described down by the picture in post #165 in this thread here:

http://www.avsforum.com/avs-vb/showt...&&#post9313612

When I was up at Brightside I wanted to measure the ANSI CR for the display, but the 9-volt battery in the lightmeter was dead and none of us had a spare. It might have been higher than we could have measured anyway though.

--Darin

Matt: Welcome. Enlighten us. Did you bring your flamesuit?

I have spent a good part of a slow day reviewing some relevant papers on the topic of what we can actually see in terms of CR as it relates to wellwatching a movie. Most of the links mentioned kinda miss the mark so I thought it better to return to the fundamentals rather then the details of human eye physiology which lose me anyways.

I downloaded these papers and unfortunately they can't be posted because I had to buy them:

1. Local luminance and contrast in natural images
Robert A. Frazor 1, Wilson S. Geisler; Vision Research 46 (2006) 1585-1598

2. The statistics of natural images
Daniel L Rudermant; Network: Computation in Neural Systems 5 (1994) 517-548.

3. Contrast Gain in the Brain ; Geoffrey M. Boynton, The Salk Institute


Summing up the main points is a difficult task at best. I usually have to read papers such as these several times to get the most out of them but there are a few points I can share.

All of these papers discuss something called contrast normalization or contrast gain control. It is the process where the eye and brain adapt to the contrast range in a given scene by finding the mean level of contrast and it occurs very rapidly, about 20 ms or so. Contrast discrimination occurs around the mean and is done partly in the visual cortex.

There is a distinction between local luminance and local contrast in a natural scene and they do not correlate. I think this is why the researchers in one of armadillos links decided to make luminance a constant because it doesn't seem to matter.

It is the difference in light levels not the light level itself that is important; this difference is contrast. Photoreceptors in the eye have a dynamic range of one to two orders of magnitude in detecting light intensity; however, do not mistake this statement to mean that the eye can only see a 100:1. That would imply that only the receptors in the eye are involved in contrast perception and it is clear to me that the brain has a big role as well and maybe what increases the perceived ratios to much higher than that, I am still unclear on this.

The Contrast Response Curve (not CSF) is the S shaped curve we are familiar with and shifts over a backdrop of a much larger contrast difference. By and large the curve stays the same as it adjusts to the mean. So the range doesn't actually increase per se it just moves around.

After reading these papers and others it is not obvious whether the detectable ratio is 130:1 or 1000:1 or somewhere in between. I'll try to tease it out of the math if possible but it is not trivial. The HVS is a very complicated thing and scientists freely admit that there is much they do not know.

CORRECTED: 2 ms to 20 ms

I can definetely refute the contrast adaptation taking place within 2 ms. Nothing in the nervous sytem occurs at that speed. A single action potential lasts 1-2 ms, that in and of itself makes it impossible for anything to adjust itself within 2 ms.

My bad, 20ms.

I might as well start this with the normal disclaimer blah blah that I'm speaking on behalf of myself and not my company. Besides my normal preference to discussing the science, I don't want the marketing guys chasing me with pitchforks for not clearing stuff with them first. So, while I'm generally going to pass on BrightSide-related questions beyond what we have whitepapers and such on, I will say that at least one light meter has made a loud popping sound and stopped working because of the display.

That experiment looks interesting. I'll have to sit down and convert stuff to candela/m^2 to try and make an intelligent guess on whether you could or not. I'll try to get back to it in a bit. You could probably do a low-tech version of the same test by just printing some checkerboard patterns and sticking one outside your window in the sun, and one inside your window in the shadows. Even if you line them up so you can see both simultaneously, you'll be able make out both checkerboards. Spot meter them, and the contrast ratio should be at least 1000:1*

* I can't do this test for at least 4 months, until Vancouver sees sun again. We're currently experiencing low dynamic range conditions out. Your mileage may vary.


Yep, I never use the Internet without one

Thanks. They were just ballparks, so in cd/m2 you could just use.

White laserpointer: Much higher than 40 cd/m2
White square: 40 cd/m2
"Black" from projector: 0.1 cd/m2
Posterboard: 0.005 cd/m2

Just want to make sure I am understanding you right. Is your position that you would be able to make out both? Would there be a change of focus required? How about an adaption time between them?

--Darin

I don't mean to patronize here, but it's slightly odd that I have to explain to you the basics of display performance. We've had those discussions in the past quite thoroughly elsewhere on the forum, and I would hope if you're going to participate in this thread that you know a thing or two... Suffice to say, light spill in a display system(including the room) reduces the simultaneous contrast ratio a display is capable of achieving (usually quantified by ANSI contrast, or similar). This spill reduces the visibility of fine detail especially near black(you can explore why on your own). The small 4% boundary difference in the pattern is just such a fine detail (relatively speaking) near black. Depending on the performance of the display (related to ANSI CR), the display may or may not be able to render your test pattern very well. This will significantly affect your ability to see the 4% square (or other near-black elements) simply because the display cannot render it. If it's not there, you can't see it. So simply stating that you can't see elements in a particular pattern has dubious meaning because it's so intrinsically tied to the display's performance.

The ANSI CR of the LCD I'm viewing is fairly high, and I can see the squares. I also expec the gamma on this display to be fairly low, being a computer monitor. I haven't measured it, but it's a fair assumption, and that too would accentuate details near black by expanding the contrast range nearer black, and that makes boundaries like these more visible. On a CRT I would not expect to see the 4% square, again because the ANSI is much lower. I haven't tested it specifically, but I've seen enough test patterns on my CRT which has an ANSI CR <100:1 to know that it's unlikely to be able to see the 4% element very well. That is because the CRT is not able to create the kinds of contrast ratios which allow that 4% boundary to be large enough in % luminance to be visible as a boundary. That is, the spill plunges the boundary below the roughly 1% JND threshold, so we cannot see it. That's why judgements based on what we can see on a pattern that is limited primarily by the display are wrong. That's why I said that this pattern has limited relevance, except insofar as it can be used to judge the performance of the display device.

Matt, do you have any supporting references to this statement in your paper:

"The eye can capture approximately 5 orders of magnitude of dynamic range effectively simultaneously."

It would be interesting to understand how this was derived and the experimental setup which produced the results.

Thanks

Heh. I thought it might raise more questions than it answered. The idea is that even though your eye is moving around the scene and the dynamic range you can see is changing, if at any point you can see the difference between both sets of black and white squares, the dynamic range you can see must at least be that great. But, it's ad hoc at best, and not something I'm going to try and build a case off of.


No, I don't have anything offhand. However, I've seen it mentioned enough places, I'm willing to accept it as fact. I'm sure this is less than you're looking for, and I'll try to hunt through the textbooks and references I have at home. I've seen charts of normalized photoreceptor response going from 0 to 1 over 5 log units. These were done by actually measuring the electrical response produced by the occular cells, I believe. How, I dunno. I've only read it second hand, but here's the book if you're motivated:

Stephen E Palmer. Vision Science: Photons to Phenomenology. The MIT Press,
Cambridge, Massachusetts, 1999.

Matt: Sorry, photoreceptors absolutely cannot possibly generate a 5000:1 response, neither in terms of amplitude (ranging between 0 and -70 mV) nor in terms of action potential frequencies (maximum 100 Hz). Actually, photoreceptors don't produce action potentials, only the retinal ganglion neurons do. BTW, there is no such thing as an ocular cell (the term refers to any cell type found in the eye; just like cerebral cell refers to any cell type found in the brain). Anyway, I would be very interested in seeing a reference or a graph. I have seen werewolves mentioned enough places that I'm willing to accept them as roaming the woods

Also, can you explain the simultaneous contrast phenomenon mentioned in the previous posts? Don't you think that the 5-squares test is quite similar to what Darin has mentioned?

I didn't mean that they produced a 5000:1 response, but that they responded over a 5 log unit range (5 orders of magnitude). Ie, they outputted some response (which the chart showed in normalized units, so I can't conclude anything about the output range) over a range of input intensities corresponding to a contrast of 10^5.

Sorry, I wasn't more specific with regards to what cells were measured because I don't know. I recall the chart had something to do with measuring the electrical response produced by some part of the eye. Photoreceptor, ganglion, or any of the web that connects the two, I can't say. I saw it at least a year ago. Not that it helps, but I also recall someone studied fish eyes a lot in the context of this problem

I think they may have been reproduced in Reinhard's High Dynamic Range Imaging book, but don't recall how much of the background was given.


I'll try. I've gotta lot of catch up reading to do in the thread.

We agree that photoreceptors (rods and cones) can capture a range of up 10^14 in total ( from a few photons to damage limit) in adaptation, and < 10^5 at a single adaptation level.

Are you suggesting that the HVS can discriminate differences in light intensities of > 200:1 at a single instance in time (visual frame)?

Are you suggesting that the ganglion cells (retinal neurons) have the ability to discriminate differences in light intensities of > 200:1 at a single instance in time (visual frame)?

Is there a typo there? Are you asking whether or not we can see differences greater than 200:1?


Without finding the actual references in question, I'm not saying anything other than some sort of charts on that topic exist.

I can't speak for HoustonHoya, but I doubt that is a typo. That is, in essence, the basic claim that started this thread.

The essence of some of the claims in the thread is that we cannot see more than a few hundred to one across a single scene. Therefore displays with simultaneous CR performance greater than say 200:1 or so are unnecessary because they already exceed our visual capabilities. There have been a variety of specific numbers suggested that follow that basic claim, 100:1, 130:1, 200:1, 250:1, 300:1. Maybe I'm missing some. But they're all a few hundred to one basically.

It turns out that the ISF is not actually stating this at all and is instead misinterpretation by some dealers who were ISF-trained, and Silver instead cited Poynton's figure of 1,000:1 simultaneous capability across a scene. This is an order of magnitude lower than the 10^5 figure from the Brightside papers which seem more accurate to me intuitively. However, despite the fact that the original post of the thread which seemed to insinuate that the ISF was unfortunately parroting this low 200:1 capability figure was actually not accurate in terms of what the ISF is saying, clearly many people have come to try to defend these low figures as correct. At least for myself, it's very difficult to take such low claims as 200:1 across a scene as the max human capability seriously, but apparently this is fairly commonly held belief. Which is why this thread is already 9-pages long. I am not sure how people are capable of driving around at night with only 200:1 capability, yet somehow...

I assumed as much, but figured it wouldn't hurt to clarify given how much attention is paid to exactness here.


10^5 is the number I'm familiar with. Perceiving contrasts greater than 200:1 or 1000:1 can be verified if you have access to a light meter. The poor man's solution is if you have access to a decent camera with a spot meter mode. Fixing the aperture and looking at the shutter speed it tells you in the highlights and the shadows should yield a significant contrast ratio. From the photography standpoint, Ansel Adam's Zone system encompasses 10 stops, or a 1024:1 contrast ratio, for film. It's not too hard to find a scene where my eyes see more than the camera in any one exposure.

On a side note, what about the argument about not needing more than a contrast ratio of 200:1 due to encoding standards? Assuming everything is calibrated right, a video system should map each intensity value to a separate JND. Now, if I increase the contrast ratio enough, I'll have more JNDs than encoded values and start getting perceivable steps in gradients and whatnot. Such things depend on all the related topics such as ambient lighting, but perhaps that's related to the confusion?

Hopefully HHF can clarify, but I think his position might be that for one scene off the screen we need more than 200:1 because we can move our eyes. But that if our eyes stayed looking at one place and the scene was static, we couldn't pick up any differences outside of about 200:1. It is possible that our eyes adjust so fast that it seems to me that I can see multiple steps at once that are well outside 200:1, but that it really isn't instant. Either way, it is easy to show in less than 5 minutes that we can pick up differences well outside 500:1 or even 1000:1 simultaneous off the screen without the scene changing (just by looking at the transitions).

There is one person who's stance has been that 8 bit video cannot do more than 219:1. He has quoted your paper and others looking for anything he can find that says "linear" or something like that to support his claim and won't do any measurements of light from displays. If he did, he would have realized how ridiculous his claims were long ago, but it is possible that he is only here to annoy people and doesn't believe his 219:1 maximum CR claim with 8 bit video with video black at 16 and reference white at 235.

The unfortunate truth is that with a dark room and no dithering, banding is a definite possibility with 8 bit video. Dithering edges can help a lot, but if you use the JNDs with 8 bit video without dithering you would get an incredibly low CR range to avoid banding altogether. I may be just saying what you already know, but if we use a 1% difference as the maximum difference between any two levels, then each level can only have a CR of 1.01:1 to the next level down. Just considering grayscale (not changing the color mix) 219 steps up only gives about a 9:1 range (1.01:1^219) to avoid all gradients larger than the JND.

--Darin

Chris,

I do see that gamma is not relevant.

In fact, the system itself puts gamma correction in at the video camera and then takes it out again at the display, by virtue of the display's (inherent or simulated) gamma. By the time the light from the display screen hits the eyeball, gamma correction is a distant memory.

So explain to me how specular highlights turn out to be reproduced on the display screen at 1/10 the relative luminance they had in the original scene, i.e., at the level of diffuse white. For this is what Poynton claims on p. 83 with, yes, no reference to gamma.

I dwell on this because if specular highlights are reproduced at a 1:10 ratio to their original relative luminance, then a daylight scene with a 1:1000 dynamic range would require only 1:100 at the display.

Thank You!

http://www.poynton.com/PDFs/GammaFAQ.pdf

http://www.displaymate.com/ShootOut_Part_3.htm


Chris I have explained this to you several times, but you are not able to grasp the concept i.e. JND, perceptual coding, and the subject matter of gamma!


PLEASE see relevant graph:
http://www.displaymate.com/ShootOut_Part_3.htm Figure 1

Have a nice day

Darin: I appreciate that you are reading and understanding what we are trying to say. There is no question that real life scenes have a high CR often exceeding 1000:1. There is also no question that we can discriminate those high CR. The real question is whether our vision accomplishes this feat simultaneously for the entire scene or whether this is achieved by by selective focusing on "subimages" of the scene and contrast adaptation occurring over time (20 ms - 20 sec) using a smaller sliding CR range of 200:1. Thus, the difference between the HVS and say photography would be that the HVS has a built-in dynamic ISO adjustment, but this takes time and is not instantaneous.

In terms of displays, I would reiterate that a small TV monitor would not need high CR whereas large projected images need higher CR than currently available if they want to be perceived more life-like. In the latter case, our HVS can take advantage of foveal movements and contrast adaptation to appreciate the higher CR presented; just like in real life scenes.