Setting contrast with Spears and Munsil disc - Page 11

Fair enough.

But I was expecting data there, I hand't thought much about which channel it would be.

Me too, I think I was expecting the distribution of colours to be even as I was thinking the cause was most likely due to scaling ringing.

To me this shift in levels points to something like a gamut mapping after the grading.

John

We're definitely seeing numerical data above 235 in both YCrCb and RGB.

What we need to identify is whether the values above 235 represent useful picture information towards white or whether its "just" noise.

Even if it is valid picture information we should also come to some sort of conclusion as to whether its of such little benefit visually as to not warrant disclosing it at the expense of contrast.

I think the only way we will do this is by identifying suspect frames and viewing them with a clip at 235 vs unclipped. This isn't just something that should be undertaken only as a frame comparisson but also "roadtested" by people running their setups at the different levels and seeing what conclusions they can come to.

To recap:

We've identified that the data in most cases contains level of some description above 235.

Is this real detail or noise?

Do the highlights visually benefit from maintaining range above 235?

Is broadcast behaving differently to BD/DVD ( is it less precisely mastered?)

Is freeing up the contrast more visually impactful than hanging on to the odd superwhite pixel?

---

If we need to consider a compromise between the two values we will at least be able to make a much more informed decision.

Many thanks to John for his analysis scripts and everyone who is producing data.

Let's not forget that clipping in YCC space is not the same as clipping in the display!

I think there is a strong argument for ALWAYS calibrating as if you expect RGB values over 235. If a clipped YCC value produces a value over 235, and you have your display aligned where red clips at 235 and D65, any increase is going to discolor and could effect your perception of the entire image.

As we've seen in this thread turning superwhite off inyou PS3 causes clipping, but we can stil see ramps up to 240+ on a per channel basis. So unless you can clip in RGB space an input RGB, you have to assume you're getting values for indvidual channel coming in at what will end up being above 235 as the display is calibrated.

Also the effects of this simply cannot be demonstrated unless we had two display next to each other calibrated for exactly this, much the same way you can't daignose contrast settings or grayscale over the internet. The file would look different on my monitor in sRGB and the camera introduces all kinds of color accuracy issues. So all we can do is discuss theory.

John,
With the RGB one could you also create an output list of the respective YCbCr values when the RGB are over the range but the YCbCr are in range?
Also, could you point me to the file in the source code where the YCbCr->RGB coefficients are defined? I would like to take a look at them...
Once again, thanks for this great tool.

The basic coefficents are in libswscale\\yuv2rgb.c
These are 2^16 * the floating colour coeffs for 16-235 yuv to 0-255 rgb.
These are their numbers and look reasonable if not exactly what I get starting from Kr Kb

I then use brightness and contrast to bring them back into 16-235 range. These are set in function sws_getContext in libswscale\\utils.c.

Also here is the quick and dirty colour matrix selector.

The code to dump the results is in libavformat\\wtwfindenc.c
Should be fairly simple code to understand, easily hackable to output whatever you want.

To build the exe you'll need an mingw/msys environment, follow the instructions here http://ffmpeg.arrozcru.org/wiki/inde...itle=Main_Page

John

I don't think Mr.D has ever suggested clipping in YCrCb space, I believe he is referring to display clipping only in his post.



I think the effects of loss of high end on an image by clipping off RGB at 235 can be well demonstrated by performing this operation in a paint application and comparing the image either side by side or a/b.

The gain of increase contrast I agree can't be A/B'ed simply but I think at the moment there is still debate as to the visibility of the content above 235.

Clearly there are those who will want "just the bits" and for them there is no question that going all the way to 255 is the right thing as there is demonstrably content there. But there isn't much and it also often looks like ringing and mpeg noise, look at the above 235 in the almost famous shots, especially the sky which is mostly flat 235 white except some very mpeg4ish blocks.


In theory then we need to calibrate up to 255, in practice for BD with somebody who perhaps needed a bit more light.... at the moment I'd be hard pressed to point to any material that would show up what they would lose by display clipping at 235 especially not the kind of shot where they would instantly change their mind. That's not to say that such a shot cannot be found...

John

Fundamentally what we need to do is be empirical about all this.

We look at imagery in RGB with a setup preserving range and the same imagery with a setup that clips at 235. If we cannot see any significant difference in the white detail between the two then we should maximise the contrast of the display as much as possible.

If we find that white detail looks flat and clipped then we need to maintain range all the way to the point that intensity variation exists on the imagery: this may be 254 it may be somewhere between 235 and 254.

I've been running with 235 as the white point on my display since last week and to be honest I like the contrast improvement on BD material.

My current display is a panasonic 42PHD8 plasma ( I do have a JVC HD1 but its currently boxed up until I get my cinema built). My primary source is an HTPC.

The plasma hardware is calibrated using Calman and an i1LT although I've also caled it with a spyder 2 and a colormunki in the past.

The HTPC as a baseline is outputting video levels rather than PC.

The plasma has an odd contrast control in that it never clips. Its essentially just some big shoulder adjustment. I set it so that it minimises crushing towards black ( yes its that weird). There is an "input level" adjustment that appears to control where the white point maps to. If I set it so that 235 is clipped on display I get some run out in the blue channel so the highest it will go without problem is 240.

However what I do is set it to 254 and use John's Upsilon Mixer to profile the display and then lut it with a custom white point of 235.

The advantage of doing it this way is that I can have multiple LUTs for different targets and punch them in and out on the fly without recalibrating the display hardware.

My current observations are: BD looks great with a clip at 235 in place. Every single time I suspect I'm seeing imagery that is possibly clipped I find when I disclose the whites to 254 that there was no perceptable image detail above 235. I'm about 99% certain that clipping at 235 for BD at least is the preferred approach. However I am still investigating this.

Broadcast is a different kettle of fish. I suspect that I'm seeing useful white detail clipped off with a 235 setup , however I have yet to verify this to my satisfaction.

Let's try to avoid jumping to one side or the other . Lets just see what the data tells us and how this relates to the displayed images. Its not a contest after all we all just want to maximise our our viewing experience.

Agreed


Agreed for the moment it looks like keeping the range may be prudent.


From what I've seen the data are fairly clear i.e. there are frames with above white.

On the empirical side with images that I've tested with I find it hard to tell if above nominal white is removed, on a few of them I can just spot a difference, but only if I know exactly where to look. The images without WTW appear artifact free. In a blind test type of setup I'm pretty sure I'd be about as good as a monkey at telling random shots apart.

John

I've grabbed alluringreality's almost famous image above.
Things to bear in mind: its been converted to png initially by alluring and I've processed it in shake with 32 bit precision then back out again.

This is with everything up to and including 235 clamped to black:



This is the original image with everything above 235 clamped had to output this as a jpg to get it uploaded:



This is a crop of the hair detail where I can see a difference as a result of the 235 clip. Checked it and the compression hasn't had too much impact in the visual difference compared with aluring's original.



The difference I'm seeing is a couple of pixels which I do think is detail rather than noise or ringing. However the visual impact is negligible in my opinion especially compared with the visual contrast advantage if you ran with a 235 white point.

It was the 3rd one that I saw the mpeg-ish blocks in the sky btw and that was using threshold in GIMP. With the girl's arm being the just noticable highlight with maybe a slight shift on the road bottom left too. But nothing objectionable.

John

this the one John?

No the other one, (sorry I can't see the originals at the moment, pesky corporate firewalls), but from memory same arm though.

John

Actually nether can I ! I can only see the two frames I posted. Never mind.

Is there any way that your WTWfinder tool could spit suspect frames out as RGB images on the fly?

What's this with corporate firewalls. I thought you were a free man not a number these days?

Maybe, ffmpeg does already have a image dump facility, on a quick test v2 does level expansion though. I'll certainly take a look.

John

Great thanks John.

Thanks.

Still a number 2 days a week until I get my head above water.

John

That's not so bad. Got to put food on the table and all that.

Update from Mr. Poynton:
http://www.spectracal.com/forum/view...hp?f=51&t=1962
http://www.spectracal.com/forum/view...&t=1953#p12486

Interesting that he says he thinks some studios do clip for whatever reason. If they are doing it higher up the chain i.e. before rescaling to HD or SD that is certainly consistent with what I'm seeing, no "real" content there only what could plausibly be scaling ringing or chroma downsampling artefacts.

A lot of the scenes (for example the almost famous ones above) contain highlights/whites that it's hard to imagine film handling so evenly, so something is pushing things to 235 at some point in the process.

John

Thanks Sotti glad to see Chuckie responding to that.

This is exactly what I would have said prior to investigating a reasonable number of transfers.

I'm still seeing nothing in BD that makes me think mapping peak white to 235 is a bad idea and the numbers we are digging up do seem to bear this out. I am 99% certain that 235 is the way to go with BD.

I'm still not entirely sure about broadcast as I've said. I'm seeing more above 235 than with BD but I'm not yet convinced its enough to worry about.

There are times when I think "thats a bit clippy looking" but I usually find that disclosing through to 254 makes no visual difference in the whites to the end result. ie there was nothing up there anyway.

Now this is usually live studio footage I see it on the most , when I'm watching high quality drama or documentary : CSI , Seven Wonders of the Solar System ( which I worked on) I'm not seeing anything above 235 even when I'm sure its going to be there.

What I see is edge features ie ringing: The almost famous example shows this...most of the region above 235 does not make sense in terms of image content. For example the region on the car bonnet you would expect the centre of the flare to be the brightest point ,which it it is... until you get above 235 where the edges start to excurse for some odd reason ...doesn't make sense, certainly not in terms of photographic image content. To be honest it does smack of mastering error to me given the nature of the content above 235.

Its some sort of filtering issue where the edge detail is banging off the 235...ringing , bad interpolation, possibly even too crunchy a grade in that region.

I'll pick up "Seven Wonders of the Solar System" on BD and have a look. I know that our content on that was delivered as "video" level material
designed to be viewed with everything below 17 clipped to black and range up to 254. In fact I had to advise another team who were working PC level (entirely CGI shots) and then panicking about what happened to their blacks on a broadcast display.

Interestingly the cms we use on our workstations never clips at 16 or 235 it always shows some range below 16 and everything up to 254. I actually drop a secondary clamp in for the blacks for viewing purposes but the work is done with realisation for the necessity of range below 16 and range above 235. If I don't get that I'll knock the shot back. I do also check shots with a hard clip 16-235 and gamma variation to confirm robustness. However it is rare for us to be responsible for the final grade...my job is mainly to provide good looking match graded imagery that does what the client wants and is robust enough to stand up to additional color correction for the final grade.

(I've never had any video work rejected by the BBC or ITV but I have worked in places where it probably would have been if I hadn't highlighted some level problems...nasty clipped holes below 17 for example)

All broadcast displays I've seen show level all the way up to 254, some show a tiny amount below 16 but I'd say its probably a function of the CRT rather than a deliberate setting.

What I want to avoid is us going over the same old ground again with this thread. For example two camps forming , 235 camp saying that there is no useful level over 235, 254 camp saying that there is level over 235 so we should keep it and here are some papers that mention excursion.

Otherwise we will never get anywhere.

Johns WTWfinder is going to be very useful in this investigation.

What we should do is continue to analyse footage and once a reasonable selection of numbers are in we'll talk about what they mean and reach some sort of consensus...whatever that may be.

Lets not look for evidence that backs up one side or the other lets just get our own evidence and see what it tells us.

And by way of experimentation try running the alternative way on your displays for a bit and see what your observations are...its not going to kill you.

Is it possible that a gamma corrected image thats strictly 16-235 when interpolated would produce overshoot above 235 especially around edges? Would the same be true if it was linearized prior to filtering?

Anything more complex than linear interpolation can and often does introduce results outside the bounds of the inputs. In my other life we go to quite some lenghts to avoid exactly this issue in 1 dimension favouring methods that don't ring.

Doing it in either gamma corrected or linear, you'll see the same issue, if anything more pronounced in linear as the normalized grey/white difference will be greater.

So it could be the case that the material was converted to video colorspace with a strict 16-235 mastering (from 2k log for example) and then the subsequent resample to HD/SD generated ringing into the range above 235?

I could probably test this when I get the time.

Then you get the actual encoding itself. I wonder if that can generate range outside 16-235 as well.

Did a quick test. Generated a 2048x1152 black and white checkerboard. clamped the levls off to 16-235. At 2k its completely clean 16-235 across the edges.

If I rescale it to 1080p I get level around the edges below 16 and above 235: as high as 249 on a brief examination. This is with a sinc filter which I'm led to believe is the most commonly used filter for downsizing.

Other filter types produce different behaviour: lanczos gets up to about 239.

A simple box filter (nearest??) doesn't produce any ringing.

I'll post images when I get a chance.


2k (cropped)



1080 sinc resize (cropped)



Range over 235

Yes this is what I thought might be happening when I saw where the WTW was. Do you remember that thread about cropping versus scaling from 2k to BD, this would be a good argument for cropping.


I'd say that's 2 parts as there is the chroma downsampling as well, depending on how this is done I can certainly see the potential for issues here. If interpolation/chroma downsampling is done in YCrCb space you certainly could get RGB > 235 spat out the other end. Since we are seeing Y outside the nominal range my guess is that it's the scaling stage that's the root of it.

The encoding could be a source of noise especially if there is a shortage of bitrate. But on BD that shouldn't be too much of an issue, there's some chance that some bits on B frames might be +/- 1 or higher on high motion scenes when you're getting close to peak bitrate. The overshoots would tend to be more regular looking though, most of what I've seen is more "organic" than that, except the right hand sky in the that 3rd almost famous shot.

John

Good point.

I created and rescaled the above images in RGB. I could convert to YCrCb prior to the rescale (including a simulated 4:2:0 downsample) but I'll need to think about how to do that meaningfully. I have no control over the coefficients : I think its smpte but cannot guarantee.

I've worked out how to do the grabs properly so.

http://downloads.upsilonsoftware.com/ffmpeg_wtwfind_3.zip

For those interested the source is available here

http://downloads.upsilonsoftware.com/ffmpeg_wtwfind_src_3.zip

What happens now is that each time it finds a frame which has a peak value or number of pixels in a channel over the highest so far it spits out a tiff file. typically that means you get more frames from the begining but it means you will always get the highest peaks and the one with the most above 235 pixels in it. Any other sensible rules? I filled up a hard disc with a rule that spat out every frame at peak

The chroma upsampling is not the best but I think the grabs do give some indication of what's there.

John

Mamma Mia - Blu-Ray (UK)

Image grab from frame with most WTW pixels.

http://downloads.upsilonsoftware.com...mia.063644.tif

John