The spectral sensitivities of the R, G, and B filters on most cameras, R, G, and B emulsion layers in most color films, and in the human visual system, all show considerable overlap of this sort. It’s not just RED.
Do the same test on some other cameras (e.g., EX1, maybe F23/35 if we can get some time on one; even an HVX200). Even buy some film, if you wanna get serious… I think you’ll find that most of the color imaging system we have to play with will show similar results.
Posted by Adam Wilt on 09/23 at 06:47 PM
Thanks for running this test, Art. I know I’ve seen a graph somewhere that shows RED’s crosstalk. I think it was the John Galt\Larry Thorpe presentation at Panavision. Demystifying Digital Camera Specifications, I believe. Link is here:
http://media.panavision.com/ScreeningRoom/Screening_Room/all_posts.html
Your graphs on the third page were particularly good at describing the problem. Thanks!
Posted by Graham Futerfas on 09/23 at 08:13 PM
Adam, I know this rarely happens, but I have to disagree with you. The RED is an extreme case.
For example, I’m looking at the same chart, shot at the same time and under the same light, by an EX3, and while there is clearly some crossover (if there was no exposure in the blue channel on the green side of the chart then the notch should drop to zero, and it only drops to about 40 units) it’s not anywhere as bad as I see on the RED.
And when I look at the EX3 image instead of the scope, the colors pop and look more saturated than the RED does. Anything containing green on the RED tungsten chart just has too much blue in it, and it’s obvious. On the EX3 that’s not the case at all.
There’s nothing colorwise in the EX3 footage that makes me afraid to shoot with it under tungsten light. The RED is another story. And I don’t think I’m the only one.
Posted by Art Adams on 09/23 at 09:53 PM
I didn’t say the RED wasn’t different than others, just that channel crosstalk is common!
If you have those same tests shot on an EX3, what do the corresponding parade displays look like? It would be most illuminating to see ‘em…
Posted by Adam Wilt on 09/24 at 12:52 AM
Hi Art,
As always an interesting and informative read. However I must point out a slight flaw in your logic.
As you rightly point out each photosite detects light but not colour, that is to say it is essentially counting photons but has no way of determining the wavelength. It then checks memory to see what the colour bias of the photosite should be.
Basically each photosite records only luminance and then applies that luminance to either R,G or B after the event. If there is a perfect filter in front of the photosite (coupled with perfect electronics and software) then in theory you end up with an exact representation of the R, G or B wavelengths that fall on each photosite.
As you have pointed out the filters on the Red are probably not perfect (nor are they on any other camera for that matter). However this is not really relevant as the photosite itself is not detecting the wavelength of light that is hitting it, it is merely counting photons.
In this case, if some amount of green (as in your diagram) is passing through the blue filter the photosite beneath should merely be counting extra photons, i.e. the blue would become more saturated (what a great way to help boost the blue channel…)
I would imagine that part of the de-mosaic process takes into account known measurements of the inconsistencies of the filters themselves and then adjusts the data accordingly (in essence, error corrects). Perhaps the new colour science in Build 20+ has adjusted the “error correction” process to allow blue to remain more saturated if to otherwise do so would allow the saturation to drop below a certain level? (Huge speculation).
However it works, if the photosites are not sensitive to wavelength the cross-talk you’re seeing must be a part of the later translation process, not of filters passing through other frequencies of light.
Posted by Stephen Webb on 09/24 at 02:11 AM
As Adam said, all cameras will have some crosstalk between channels—particularly single sensor cameras that rely on dyed filters as opposed to dichroic filters—this is why every camera applies some sort of color correction matrix to the images at some point in the chain (which in fact is doing exactly what you did with the channel mixer filter.)
This is not necessarily a bad thing in and of itself—it’s easier to get a sharper image out of a bayer image with a fair amount of cross talk, for instance, but it will definitely lead to more chroma noise. And at the end of the day if not enough photons of particular wavelength are reaching the sensor (or I should say being “sensed” by the sensor), you’ve got a problem no matter what, which is really the root cause of the Red’s poorer performance under Tungsten light compared to daylight.
Posted by Jesse Rosen on 09/24 at 08:03 AM
Hi Adam-
The EX3 waveform now sits at the bottom of page 3. Let us know what you see…
Posted by Art Adams on 09/24 at 10:50 AM
Hi Stephen-
I think I see a flaw in your logic. 
You state that a little green passing through the blue filter would make blues more saturated, but that’s not the case because green will only be present in an object which contains green already. Any object that contains enough green to trigger this response is not going to be pure enough blue to become a more saturated blue; it’s going to be blue/green, or cyan.
Indeed, if you look at the charts in this article and the color comparison in my previous article, you’ll see that cyan is one of the RED’s strongest colors, if not its strongest—precisely because it consists of both blue and green. The green channel, which is strong under both tungsten and daylight, contributes, as does the blue channel, which contributes more than it should under tungsten light.
This may be one of the RED’s stronger points, as most of the prettiest blues in nature contain a fair bit of green. As I’ve said before, the RED isn’t accurate—but it sure is pretty.
Posted by Art Adams on 09/24 at 10:56 AM
Hi Art
My point was that the photosite itself doesn’t detect colour, so it can’t be contaminated by green light. How would extra light of a particular wavelength effect the photosite if it can’t distinguish colour?
Posted by Stephen Webb on 09/24 at 10:59 AM
Hi Jesse-
I get what you’re saying, and perhaps I wasn’t clear in the article: yes, all cameras crosstalk to some extent, but the RED under tungsten light does it excessively—and most importantly, noticeably. This may be one of the reasons why people are less happy with the RED under tungsten light.
I think you’re saying that the issue with the blue channel under tungsten may be a result of matrixing, but considering that the matrix would have to change based on the color temperature of light that you’re shooting under, shouldn’t this test have revealed that? As best I can tell, the only way the RED, or RedAlert or RedCine, can determine the color temp of light is through the white balance meta data, so by setting the white balance to one color temp but shooting another very different color temp should have revealed that matrixing, no? The matrix should have been applied where it wasn’t needed.
Instead what I believe I saw was the RED simply responding to the light in front of it, without any matrixing/channel mixing to boost the blue channel under tunsgten light.
Posted by Art Adams on 09/24 at 11:06 AM
Stephen- but it can be contaminated, precisely because the photosite can’t see color!
If we’re shooting a pure green object, here’s what happens:
Red photosites: see very little, don’t really respond, no significant red is added to the color mix.
Green photosites: see a lot, respond strongly, trigger the addition of a lot of green to the color mix.
Blue photosites: should, like red, see very little, because there’s no blue in pure green—but instead do detect some illumination due to the blue filter allowing some green to pass, which triggers the addition of blue to the color mix even though the object contains no blue.
The result is that a pure green, under tungsten light, will be muddy because the blue photosites are detecting the green as being blue, and are triggering the addition of blue.
By the way, I may be starting a series of tests where I look at what kind of crosstalk is present in different cameras. Stay tuned…
Posted by Art Adams on 09/24 at 11:14 AM
>the RED isn’t accurate—but it sure is pretty.
Hi Art, you keep saying this. Would you mind elaborating? What do you find particularly pretty about these images? I’ve been seeing a lot of inaccurate, sallow, ruddy skin tones on recent RED-shot movies. I don’t find that pretty. While I know that desaturation is a common part of today’s style, on my current romantic comedy movie, I would like the skin tones to look natural and beautiful. I’m concerned I’m going to struggle with this, and probably have to employ significant secondary color correction.
Posted by Graham Futerfas on 09/24 at 11:15 AM
Right, with you now.
Posted by Stephen Webb on 09/24 at 11:21 AM
So then, if my assumption that this would normally be measurable and error-corrected, does it follow that maybe that is now not happening to the same extent? Have you thought to test an older build (17 for example) and see what that is doing?
Posted by Stephen Webb on 09/24 at 11:24 AM
I’ve only shot one project with the RED under tungsten light, and even then we did a quick test with an 80D and saw that the filter helped so much that we used it for the entire project. I have to admit that I have very little 3200k experience with the RED—intentionally.
Under daylight, though, I’ve gotten some good results: flesh tone feels good, and playing with the gammas makes the RED feel like a camera with more latitude than it has: boosting the mid-tones on a naturally-lit scene gives a real feel of where the surfaces in a room are, and what they’re reflecting.
I guess what I really like about it is that it’s not an F900 or Varicam, which were the mainstays of my market. Both are good-looking cameras but I’m tired of baking in a look on-set. I like having a camera that HAS to be color corrected, because the opportunity for creating interesting looks on a fast schedule grows exponentially. No one is going to take a lot of risks with looks when you’re recording to 8-bit tape and painting in the field, but they’ll go crazy with a colorist and 12-bit raw in post.
I’ve got a director who loves the RED because we don’t have to stop and paint it, and I love that I can shoot fast without worrying too much that I’m not getting enough time to finesse the look.
If you’re shooting a romantic comedy under tungsten light, yeah, I think you’re hosed. I’d recommend sneaking an 80D in, but then you’re boosting your lighting package significantly beyond the RED’s normal needs. As you know, it’s not a fast camera.
I’ve gotten some very pretty looks out of it, but you wouldn’t want to use it side-by-side with an F35, and F23 or a Varicam 3700. The differences would be startling.
What are you rating it at on your feature?
Posted by Art Adams on 09/24 at 11:33 AM
Hi Stephen-
Unfortunately I’ve upgraded my RedAlert and RedCine and don’t have older copies of either. My understanding is that the new color science is built into both, so I wouldn’t know for sure what I was seeing if I opened up old Build 17 footage in either one.
Posted by Art Adams on 09/24 at 11:35 AM
>What are you rating it at on your feature?
320 setting on the camera, 160 on my light meter under tungsten. Haven’t worked out daylight exposure yet, it might be a little faster.
Posted by Graham Futerfas on 09/24 at 02:38 PM
Hi Art,
I don’t think the problem is caused by the matrix. You see different results under tungsten and daylight because the spectral content of the light is different, not because there is a different matrix used (apparently). Note that some raw converters will apply different matrices for different color temperatures (the D21 has matrix settings for this). See also the DNG spec which specifies how compliant DNG converters should interpolate a matrix given specific matrices for two different illuminants. In the case of Red it should be possible to get color response under tungsten closer to what you get under daylight using a different matrix, but it would probably result in more noise.
Posted by Jesse Rosen on 09/24 at 07:57 PM
“You see different results under tungsten and daylight because the spectral content of the light is different, not because there is a different matrix used (apparently).”
Yes. That’s the point of the article.
Posted by Art Adams on 09/25 at 08:38 AM
Hi Art,
For the record, you can download old versions of Red’s tools for use with old footage—in case you really feel like testing the older stuff:
http://www.red.com/support/release_history/3
I haven’t installed an out-of-date tool before, so buyer beware.
Cheers,
ryan
Posted by .(JavaScript must be enabled to view this email address) on 09/25 at 09:12 AM
Hey Art,
Sorry if I came across as disagreeable in some way—I certainly appreciate the tests you’ve done and think they are quite useful. Plus I think I have a tendency to be unclear about the specific point that I’m responding to, which in this case was that the Red seems to have more crosstalk under Tungsten light than under Daylight, which is physically impossible since the spectral response of the RGB filters on the photosites doesn’t change—what changes is the spectral content of the light.
I’m sure you fully understand this, but I was just trying to clarify what I thought was unclear.
Posted by Jesse Rosen on 09/25 at 09:40 AM
