Reasons to go for C300 over RED Scarlet X?

[Emmanuel Plakiotis]

Without being an expert, I think that the beyer filter addressed the technology inadequeacies of a bygone area. The most modern sensors (F65, C300) are not beyer and I don't think this trend is going to fade.

Many have point out that this new sensor readout "impersonates" a 3 chip camcorder. I wonder, isn't more accurate to say a 4 chip camcorder?

[Tim Le]

The C300's sensor still has a classic Bayer pattern. It's just not doing any de-bayering to reconstruct the image. There are exactly 1920x1080 red, 1920 x 1080 blue and effectively 2X 1920 x 1080 green photosites. There are twice the number of green photosites because green wavelengths account for most of the visual information.

The chip has four parallel readouts to readout these photosites directly: one each for red, blue, green1 and green2. The two greens are then combined so you end up with 4:4:4 color sampling, much like a 3 chip camcorder. At least that's the way I understand it.

[Chris Hurd]

Originally Posted by Emmanuel Plakiotis

Many have point out that this new sensor readout "impersonates" a 3 chip camcorder. I wonder, isn't more accurate to say a 4 chip camcorder?

Sorry but no, that is not accurate. It impersonates a 3-chip design.

A three-chip camcorder samples green the same way as the C300 does. In a three-chip prism block, 25% of the incoming light goes to red, 25% to blue and 50% to green. Our eyes are more sensitive to the color green, and the green channel carries all of the luminance (brightness) info.

[David Heath]

Originally Posted by Chris Hurd

A three-chip camcorder samples green the same way as the C300 does. In a three-chip prism block, 25% of the incoming light goes to red, 25% to blue and 50% to green.

I have to disagree. In a 3 chip design, then all the red, all the green, and all the blue light get sent to their respective chips by the beamsplitter. So for each location, the relevant photosites get to respond to all the light corresponding to their filtered colour. Hence 100% of red light to the red channel, 100% of green light to the green channel etc.

With a Bayer filter, the individual photosite colour filters perform the spectrum separation, so three adjacent photosites (R,G, and B) are all that's really needed in principle. The second green photosite will just have the main effect of improving the noise figure in the green channel (and also improving spatial characteristics), and since that is what luminance is mostly derived from, that's no bad thing.

[Don Miller]

There must be loss of photons in a beam splitter. Doing RGBG on a single sensor maintains the bayer advantage of 50% green, while avoiding the cost of a beam splitter. In the images released so far the Canon doesn't seem to have any color artifacting issues at all as far as I can see. So perhaps it's now possible to build very clean color filters onto CMOS now. Certainly the large sensor size, compared to 1/3, must help in construction. I'm curious how anti-aliasing is handled in the C300.

The C300 has more green sensors than the F3 has total sensors. That should make for more real world detail. The green readouts may be combined on chip so the the processor sees each pair as a single value. My understanding is that CMOS, unlike CCD, can include on-sensor functionality.
I assume Canon and Sony can build more sophisticated sensors than Red or Arri can source. It's a shame we can't see something close to a "raw" readout from the C300 sensor.
I expect with Red too much credit is given to "the brain" and not enough to the image processing that runs on the PC. In many ways Red and Canon have inverse strengths.
It's likely Scarlet output will be indistinguishable from Epic. It would probably cost Red more to cheapen up Scarlet than to build it with Epic engineering. They have plenty to do making their products the best they can be rather than engineering less image quality into Scarlet.
And besides, what does "the brain" do? In a Red after the A/D conversion it only runs a compression scheme and dumps the bits out. None of that messing business of actually making a video file. Perhaps the monitor portion of a Scarlet won't be as good or sophisticated as an Epic. But it doesn't seem there's a lot of opportunity for output quality to vary between the low and high models.
Epic will now only get the very best of the sensors. But they aren't going to be putting junk into Scarlet. They would hurt their brand doing that.

[Tim Le]

Originally Posted by Don Miller

It's likely Scarlet output will be indistinguishable from Epic. It would probably cost Red more to cheapen up Scarlet than to build it with Epic engineering. They have plenty to do making their products the best they can be rather than engineering less image quality into Scarlet.
And besides, what does "the brain" do? In a Red after the A/D conversion it only runs a compression scheme and dumps the bits out. None of that messing business of actually making a video file. Perhaps the monitor portion of a Scarlet won't be as good or sophisticated as an Epic. But it doesn't seem there's a lot of opportunity for output quality to vary between the low and high models.
Epic will now only get the very best of the sensors. But they aren't going to be putting junk into Scarlet. They would hurt their brand doing that.

I agree. Since RED cameras are just outputting and compressing raw data, the results should theoretically be identical. Graeme Nattress over at Reduser has said the image sensor is identical in Epic and Scarlet. They're not binning those chips--it's just the ASICs that are being binned. The consequences seem to be pretty much what they stated: there are limits in resolution and frame rates. There is one new limitation: HDMI and HDSDI will not work simultaneously on Scarlet. Epic has the same limitation now, but it's a firmware issue that might be resolve early next year. But it sounds like with Scarlet, it's a processor limitation, which makes sense.

[David Heath]

Originally Posted by Don Miller

There must be loss of photons in a beam splitter.

Theoretically, no. (And practically, to a first approximation, also no.) If a photon of wavelength corresponding to "red" enters, it goes down the route to the sensor dedicated to the red channel. Same for green, blue.

With any single sensor device, if a photon of "red" light hits a red photosite it passes through the filter and registers, if it hits a blue or green photosite, it gets absorbed in the dye and hence is lost for photographic purposes.

Hence why a 3 chip device is inherently more sensitive than any current single chip one of the same size format.

In the images released so far the Canon doesn't seem to have any color artifacting issues at all as far as I can see. So perhaps it's now possible to build very clean color filters onto CMOS now. ........... I'm curious how anti-aliasing is handled in the C300.

I wouldn't expect colour artifacting as the photosites for each output pixel are so close together. And what we don't know is what low pass filter Canon may use.

The C300 has more green sensors than the F3 has total sensors. That should make for more real world detail.

Not necessarily, it depends on how they get dealt with. If corresponding green values just get added together, that will well impact on sensitivity, but not detail. And remember what is important is luminance detail. Because of the formula for deriving luminance (Y=0.3R+0.69G+0.11B, if I remember correctly :-)), that's why it's true to say that the green channel is most important. But most is not the same as all.

The direct read out system used as in the C300 ignores spatial differences between corresponding R,G,B photosites. In true deBayering, those spatial differences are not ignored - and blue, red photosites also correspond to luminance resolution. Not as much as green, but......

And that's why the more you go into it, the more difficult it becomes to think in terms of the headline numbers. I can't pretend to understand the most subtle points, but "more green sensors than the F3 has total sensors" is not necessarily true.

Maybe - maybe not.

[Alister Chapman]

Sensor theory is one thing, real world performance is another. Prism designs are ver efficient at splitting the light into the R, G and B wavelengths. There is very little loss and very little leakage between colours. But the light has to pass through a very thick piece of glass and this then causes issues of it's own (flare being just one). Bayer is clever because it relies on the fact that most real world scenes don't contain discreet primary colours and 2x2 CFA is clever because it does not need to be de-bayered. Each system has advantages and each has disadvantages. For example bayer is known to alias in some colour frequencies due to sub sampling and CFA may have cross colour issues as there may be little compensation for the overlaps in the colour filters.

In the end, I'm sure all of the current crop of cameras will produce great images. The skill will be in exploiting the strengths of the overall package. Some have simpler workflows, some have better dynamic range but maybe at the expense of a more complex workflow. Thus the "package" becomes more important than just the sensor.

[Don Miller]

Originally Posted by David Heath

...............

Not necessarily, it depends on how they get dealt with. If corresponding green values just get added together, that will well impact on sensitivity, but not detail. And remember what is important is luminance detail. Because of the formula for deriving luminance (Y=0.3R+0.69G+0.11B, if I remember correctly :-)), that's why it's true to say that the green channel is most important. But most is not the same as all.

The direct read out system used as in the C300 ignores spatial differences between corresponding R,G,B photosites. In true deBayering, those spatial differences are not ignored - and blue, red photosites also correspond to luminance resolution. Not as much as green, but......

And that's why the more you go into it, the more difficult it becomes to think in terms of the headline numbers. I can't pretend to understand the most subtle points, but "more green sensors than the F3 has total sensors" is not necessarily true.

Maybe - maybe not.

We can count photosites. The C300 has many more green. But you're right in that the total area (and sensitivity) should be equal between the two cameras. However the C300 photosites are perfectly distributed. Calculating spatial differences makes up some of that difference in the F3, but still has to have lower true resolution. You can't measure much less and get back to the same resolution.

Debayering helps with a lot of bad pixels. So I think the iphone will debayer for many more generations. I think pro video debayers because that was historically the way to make less expensive cameras. It's remarkable camera companies have been able to do with post capture processing. But all other things being, starting with an image that resembles reality has to be better than what comes off a traditional bayer sensor that is less than 4x the resolution of the final output.

I think the F3 debayers because that is what fits in Sony's low mid tier product line. I think Red debayers because that's the technology they could buy. I think the C300 doesn't debayer because they are now technically capable of doing it right, and don't have Sony's product cannibalization problem.

Of course, as a large corporation, Canon had to mess it up by squeezing it down to 50 mbs and 8 bit. Which for me makes the Scarlet more interesting. Maybe shooting 4K for 1080p at a somewhat higher than normal compression. I just need to get over owning 10-12 batteries.

I have no illusion that Epic or Scarlet will be meaningfully upgradable. The electronics aren't good enough. These are 3 year purchases. But that said I do agree with Alister that all of these products are likely very good. We've reached something of a golden age in lower cost high end DV. If all TV for the rest of the decade was shot with the F3 no viewer would request higher quality. I'm not even sure that anyone cares if projected video is more than excellent 1080p. My friends and family usually don't see a difference between cable SD and HD.

[Peter Moretti]

FWICT, for all intents and purposes, it's a quad-HD Bayer pattern sensor (photosite color ratio of 1R:1:B:2G).

It's not Debayered b/c the multiple readout allows for a full HD image to be reconstructed by simply combining the 4 photosites to one pixel. The F35 does a similar thing, only using a vertical stripe pattern and it doesn't sample G twice.

What's left out in all this is that this is a very inefficient usage of a photosites and pixels. A good Debayering algorithm would increase resolution by quite a bit. But then you'd need a higher than 1080P frame size for the recorded image.

[David Heath]

Originally Posted by Don Miller

We can count photosites. The C300 has many more green. ........ You can't measure much less and get back to the same resolution.

Ah, but you're not taking into account the colour element of the subject. It helps to think of "input photosites" and "output pixels". And an output pixel has three values - one for luminance, and two for the two colour difference signals.

It's pretty easy to work out (to a first approximation) what will happen in the C300 case. Four "input photosites" go directly to make up one "output pixel" - end of story. Luminance resolution will be 1920x1080 - same as from a 3-chip system.

With deBayering it's different. If you deBayered the C300 chip, you firstly go to a 3840x2160 "output matrix" - each pixel having the three values for luminance and two chrominance. Here, for luminance, each "output pixel" will have the luminance value of it's corresponding photosite - plus a percentage of the luminance of neighbouring pixels. The clever part is how the weighting gets done. That means a sort of averaging - which will mean the output luminance resolution must fall short of 3840x2160. But it will be far better than 1920x1080. The chrominance values for each "output pixel" must also be calculated - and it's easy to work out that this must be coarser than for luminance - hence a function of deBayering is better luminance resolution than chrominance. Just like the human eye!

So the next question is, for a chip that's going to be deBayered how big does it have to be to match one read out in the same way as for the C300? For luminance, the answer is "about 25%" in each direction, so nominally about 2400x1350. Now isn't that strange!? Very close to what the F3 actually is! Clever people these engineers - put the right numbers into the number cruncher, and it's pretty obvious what they will say should be done!

It's why I quite agree with Alister when he predicts that for resolution at any rate, there won't be a lot to choose between the cameras. I do wait to see what the aliasing will be like for the C300 - especially for chroma, and especially out of band. I'm not going to try and predict that one..... :-)

[Emmanuel Plakiotis]

Originally Posted by Peter Moretti

What's left out in all this is that this is a very inefficient usage of a photosites and pixels. A good Debayering algorithm would increase resolution by quite a bit. But then you'd need a higher than 1080P frame size for the recorded image.

It might be inefficient resolution wise, but in every other respect is optimum. From what I've read so far it's not a bad decision. Which brings back the haunted question. Why Canon took so much pain to create a superb color image and doesn't allow the end user to take full advantage of it...

[Tim Le]

Originally Posted by Emmanuel Plakiotis

It might be inefficient resolution wise, but in every other respect is optimum. From what I've read so far it's not a bad decision. Which brings back the haunted question. Why Canon took so much pain to create a superb color image and doesn't allow the end user to take full advantage of it...

I agree, Canon's method seem to be optimized for an HD signal and the C300 is only an HD camera. De-Bayering has disdvantages too, such as reconstruction errors and computation power necessary to do the de-Bayering. Part of the reason why the C300 uses less power is because it doesn't need to de-Bayer. Less power means less heat so the cooling system can run silently. Also, the C300's DIGIC DV III chip is expecting direct color readouts like the 3-chip sensors in the XF305 and the C300's sensor emulates this. It really was the best, most elegant solution for Canon, IMO.

Emmanuel, I think you're referring to the 8-bit limitations? The reason is just a practical one: Canon did not have a DSP chip that could handle a 10-bit baseband at the time the C300 was developed. All they had was the DIGIC DV III. However, this chip does process the sensor data at 12 and 13 bits before being conformed to 8-bits. We have to remember Canon isn't in the same situation as Sony or Panasonic, who have an established line of CineAlta and Varicam cameras that have the 10 bit infrastructure already developed.

[Peter Moretti]

Another reason for 8-bit is that there is a trade-off in terms of resolution quality and color quality at work. FWIU, at low bit depths (and 50Mbps is probably on the border of this characterization), it's better to use those pixels for resolution than for color depth.

Now it's true that when DR is large, then 8 bits can lead to more color banding. But even XDCAM-EX, which is 35Mbps 4:2:0 8-bit can accurately record the expanded range that the Hypergammas provide. Now it might crap out w/ S-Log or whatever the F65 spits out.

But I don't believe that C-Log is as aggressive a log curve as is S-Log, since Vincent Laforet and his editor said that C-Log looks good w/o using a LUT or adding a color correction layer on top of it.

[Don Miller]

If 8 million sensors is too much, Canon could have gone with a less dense sensor and the XF100 chip set. Perhaps we'll get that in a C100.
It is interesting that in the three camera, C300, F3 and Scarlet, we have the three current better sensor designs for producing a 1080p file. We should be able to learn something from that spectrum of products.