Monitoring for chroma clipping

H

Hywel

Well-known member
Hi All,

Technical question. My understanding of digital acquisition is that one wants to "shoot to the right" (in still camera parlance). That is, shoot the highest exposure you can without clipping. Shooting in RAW for stills makes this easy- just check the histograms and make sure nothing is bunched up at the top end. One should do this for the red, green and blue channels separately, and also double check the combined histogram just in case. One can often end up under- or (more usually) over-exposing compared with where one might want the zones to end up in the final image, in order to capture a clean "digital negative". It is always better to get the maximum number of signal electrons on the sensor without filling any sensor wells (=clipping) and down-scale later according to the desired gamma- this always give one optimal signal-to-noise.

When shooting in JPEG mode, my experience is that one wants to get closer to the "intended" exposure, because you are applying a strong perceptual lossy compression scheme (and losing bits of resolution, too). So it is not wise to over expose by a couple of stops and turn it down in post because you might end up devoting some of your precious 8 bits and data to detail you really won't care about in the final image. There's no sense capturing lots of (noise ridden) detail in deep blacks if you know they are going to be crushed in the final output.

But it is still very important to make sure none of the individual RGB channels clip, because if they do, you've lost colour information. So while you light and expose for the final image, you still may want to turn the wick down on the saturation just enough to make sure that a bright red dress isn't clipping badly in the red channel. You want to get a nice full range of tones, so you don't want to turn the saturation down too much either, but you must avoid clipping.


Presumably, one wants to work on largely the same principle for digital video. The problem is that the readout of RGB is done way higher up the acquisition and processing chain in a video camera than it is in a stills camera, even in a stills camera shooting JPEG, because the RGB signals from the individual colour components of the Bayer pattern on the sensor are combined into Luma and Chrominance signals before you even get a chance to monitor them.

So... I know how to avoid luma clipping- watch for signals above 100% IRE. And the AF101 has several tools provided to do just that, with multiple Zebra settings and a waveform monitor.

What I don't know how to avoid is chroma clipping. Suppose I'm shooting a girl who is wearing a bright red satin dress. The luma component of the signal from the dress might be sensible, but the individual red channel pixels might be clipped to hell. How do I spot that on set? I'm not familiar enough with the vectorscope to know if that will show this up, and if it does, how to spot it.

Do I just look for chroma values outside the ring of target boxes?
Can I use colour bars to get "pure" colours and calibrate for that in some way?
Is there a way of seeing what's going on further up the acquisition chain and checking that the individual RGB channels aren't clipping straight off the chip?

Cheers, Hywel.
 
You almost want an RGB Parade instead of a waveform.

I've always thought of shooting my HVX as shooting transparency film where I need to get the exposure pretty much spot-on, with maybe a half stop sensible latitude each side before clipping.

And if the red clips first as in your example, you're always going to have a little Red/Cyan 'tidemark' around the clipped area as you throw your grade around, as Green/Blue still have an unclipped profile.

I'd love to be able to design a lookup curve in Nuke or Photoshop, and to be able to export it to the camera and have as my personal shooting profile/mode instead of Cine-V or whichever default flavour you like. To be able to get a better, personal Cine-D would be good, that same wide lattitude, but rolling the whites of more, easing up to zero instead of the current linear clip.
Guess you can do that on RedMX?
 
Last edited:
I shot something like this not to long ago on the hmc 150 and what I did was use a different scene file .
Shot most of the dresses with Cine V and when certain satin colors cameup I went to Cine- D which to my eye was WYSIWYG !
Once broadcasted there was no clipping what so ever ....Highlight clipping can be ugly sometimes but chroma clipping to me is ugly all the time.
Ofcourse this is a totally different situation( maybe) but certain colors dont show up right with certain Matrixes.
 
Hywel said:
I'm not familiar enough with the vectorscope to know if that will show this up, and if it does, how to spot it.
Click to expand...

With a vectorscope - there will be 6 boxes - corresponding to the usual RGB - plus cyan, yellow and magenta. A good rule of thumb would be to draw imaginary lines from box to box - creating a hexagonal shape. If the signal in the vectorscope crosses those imaginary lines there's a good chance of clipping that color.
 
Last edited:
Here's an example, a fairly subtle one but it shows what I am worried about.

Redclip1.jpg

This was shot on location in evening light, on a Canon dSLR. I noticed when I got the shot into FCP that there's a hint of chroma clipping on the skin tones, but there's no sign of it on either the waveform or the vectorscope.

I only see it on the red channel in the RGB parade. What worries me is that looking at the scopes, you'd think you were exposed safely. But looking at the red channel shows you are already into >100% which is certainly into the danger zone. And because the luma and chroma information was actually built from the RGB signals from the sensor, some nasty surprises may be hiding in that red channel.

Indeed, from the look of the footage our poor damsel is not just in distress from the ropes, but also from her skin tones starting to do strange orangey-greeny-yellowy things. This is the classic Canon-stills-JPEG sign of red channel clipping and makes it hard to rescue in the grade because you're already starting to lose information- even though a glance at the vectorscope and waveform trace would not have rung warning bells for me if I'd seen them on set.

So I'm concerned that despite the HUGELY welcome ability to show these monitors live on the AF101, there's still a danger of significant clipping lurking in the footage because it has already clipped at the RGB stage.

Is there anything one can do to spot this? On the Canons it has become our practice to take stills at the start of each take and quickly check the RGB histograms- is that possible (and useful?) on the AF101?

Cheers, Hywel.
 
Thank you for bringing this to attention. Very interesting.

I'm not sure it is possible to see separate RGB waveforms on the AF100. I thought the waveform was just luma.


Do I see correctly in your vectorscope a red line spike that is clipping? Could one have used that?
 
Unfortunately not, that line is the reference line for skin tones to show where they ought to lie for average Caucasian skin tone.

I don't think there are actually any points from the picture that far out (it is easier to see on my original screen than on the web-resized JPEG screen grab).

The most saturated colours on the vectorscope are just under 40% and the highest values on the luminance are at 95% - and yet the red channel is clearly well into the danger zone.

This is the issue with going from RGB to luma/chroma without being able to monitor the source RGB channels directly.

Mathematically, you can demonstrate the problem as follows. Suppose you decide to transform from RGB to an arbitrary encoding DEF which I'll define as:
D = (R + G + B)/3
E = (R - G + B)/2
F = (R + G - B)/2

You can easily recover the RGB values:
R = E + F
G = (3D - 2E)/2
B = (3D - 2F)/2

But now let's see what happens to our skin tone. Let's choose a scale of 0 to 100, with anything at a value of 100 being clipped. Say our just-clipping skin tone has R=100, G= 90, B= 80. What do we get for the transformed colours?

D = (100 + 90 + 80)/3 = 90
E = (100 - 90 + 80)/2 = 45
F = (100 + 90 -80)/2 = 55

See what's happened? Although we KNOW we have clipping in the red channel, it doesn't show up in the transformed colour space D, E, F at all. We note D (which I've constructed to be vaguely like luma) is a bit high but at a value of 90 it doesn't look dangerous, and E and F are a very long way away from clipping at 45 and 55 respectively. Nonetheless, just because we have done a colour transformation, we haven't rescued the image.

The clipping is still there- the transformation has just hidden it from us. We still have no detail left in the red channel at all. The physical acquisition of the signal at the very first stage has saturated- the wells are full on the red pixels, and nothing we do downstream can possibly recover from that because the clipping is a physical property of the sensor. The only way we can recover is ON SET to turn the exposure down so those original red pixels do not saturate. There's NOTHING we can do in post to recover (although we can probably fudge it a bit to make it look less ugly).

So... even though the skin tones are clipped, viewing the information in a transformed colour space has made that fact a hell of a lot less obvious than it should be. Going from RGB to luma and chroma is actually a bad idea without monitoring the original RGB channels to watch for clipping.

Cheers, Hywel.
 
Last edited:
I want to cautiously ( with trepidation) enter this discussion and respectfully say:

I wonder if you are over thinking and over training in this matter?

I have never ever thought about or been told (by technical directors, colorists, fx artists, compositers..) that anything I have ever shot is Chroma Clipped.

I can not even imagine how a skin tone could ever achieve a level of color saturation that would or could over-drive chroma in a camera. (blue lady in Avatar or guy in The Hulk when roused maybe exceptions)

I have never heard any other DP even bring it up or discuss it.
 
Shooter said:
I want to cautiously ( with trepidation) enter this discussion and respectfully say:

I wonder if you are over thinking and over training in this matter?

I have never ever thought about or been told (by technical directors, colorists, fx artists, compositers..) that anything I have ever shot is Chroma Clipped.

I can not even imagine how a skin tone could ever achieve a level of color saturation that would or could over-drive chroma in a camera. (blue lady in Avatar or guy in The Hulk when roused maybe exceptions)

I have never heard any other DP even bring it up or discuss it.
Click to expand...

+1
 
Shooter said:
I want to cautiously ( with trepidation) enter this discussion and respectfully say:

I wonder if you are over thinking and over training in this matter?

I have never ever thought about or been told (by technical directors, colorists, fx artists, compositers..) that anything I have ever shot is Chroma Clipped.

I can not even imagine how a skin tone could ever achieve a level of color saturation that would or could over-drive chroma in a camera. (blue lady in Avatar or guy in The Hulk when roused maybe exceptions)

I have never heard any other DP even bring it up or discuss it.
Click to expand...

I think the one factor that we are not looking at is that as you start to expose images in to the 90's and 100's you lose saturation. The RGB Parade shows the luma values for each specific color channel. But it does not tell you how much Chroma or saturation is in that channel. Which is why if you Clip the red channel on the Parade you dont get a deep red color cast but light red color cast. So over exposing or even clipping your images will rarely lead to clipping of the chroma. Sure if you over expose or clip a color channel you may have a colored halo around the subject or have a difficult time bringing the detail back to the image, but this is because of a loss luma information in the specific color channels.

This makes me think of another point about exposure. The one thing i have noticed is that just because you've have not clipped your image does not mean that there is useful information in highlights of the image. Areas of the image that are in the 90's and especially the upper 90's are far more difficult to bring back to lower levels. For instance, if skin tones get in to the 90's and especially the upper 90's you start to lose the texture in the skin and the codec starts to blend the tone together. (BTW I never grade a source higher than 4:2:0. So these comparisons of skin tone and exposure deal specifically to highly compressed codecs like AVCHD and EX-Cam. I'm sure acquiring in less compressed codec and have less chroma sub-sampling will give me more detail in the highlights, but as it stands what i've said hold true for the formats that those in the general af100 community will be shooting.)

Ok its two am. and i want to get to bed so i can wake up early and play starcraft and hopefully get my parts in to build my arcade machine.

Its also good to see other with an insanely technical mind.
 
Last edited:
Shooter said:
Sorry.

Read that three times and did not comprehend any of it.
Click to expand...

Sorry dude, I kinda have that issue in explaining what in the world is going through my brain. But basically luma and saturation have a slightly inverse relationship. The brighter the image the lower the saturation. Which is why if you clip a color channel you get a light color cast instead of a heavy one. If all three channels are clipped equally you would get pure white at the clip point.
 
You must log in or register to reply here.
Back
Top