Enter the Dragon - 20-Stops of Dynamic-Range

Starcentral: I think you're confusing what RED Raw is and what LOG is. RAW is just the raw data that gets pumped out of the sensor. In the case of RED, this raw data is compressed to some ratio.

LOG is a gamma curve. Whether its Redfilmlog or Arri Log C or Sony S-Slog: these log curves are just another way of interpreting the RAW data for representation on visual sources. 16 bits should be quite fine for 15-16 stops of dynamic range.
 
Raw is data off the sensor yes, but it can be stored in a linear encoded format or log based format. Currently with RED, sensor data is stored in a linear encoded format as opposed to log encoded format. No one is talking about data compression.

For this amount of dynamic range 16-bit linear will not be enough, but Log would address the issue.
 
It's not about the format though. You can, in practice, assign as many stops to as many levels as you like within a given bit depth, but the sensor itself is linear; there is no such thing as a logarithmic sensor as far as I'm aware. As it is with today's technology, every additional stop of light a sensor can capture is represented by the most significant bit, under-expose by 6 stops and your 14-bit quantization becomes an 8-bit one. The bit depth of the sensor is directly proportional to how many stops of DR it can record, regardless of how those bits are stored, be it linear or log.

Considering 16-bit ADC is only starting to inch it's way into relatively mainstream imaging technology, I'd be surprised if RED have one that can quantize 20 stops worth of levels; maybe it's another RED magic trick and they have 24-bit tech under their sleeves?
 
Last edited:
It is all about format. Irregardless of bit-depth pick and compare linear, log, hyper gamma curve, and REC709 and for each "format" there will be differences yet the sensor is the same. Some will give you less apparent dynamic range, while some will give you substantially more.

Yes in practice you can assign as many stops as you like in any given bit-depth but do you not consider the overall steps of gradation in-between stops required to produce smooth tonal information? More stops means you need higher bit-depth in order to keep the gradations in-between stops smooth.

Remember with linear encoding you need even more bit-depth for smoother gradation at the lower end since per stop there becomes insufficient levels to describe the dark tones - and excessive levels to describe the bright tones.
 
I don't think young people or new people really understand where LOG comes from. LOG was first developed as part of the cineon system digital spec by Kodak for film negative scanning. That "logarithmic" curve has been adapted and tweaked, faithfully or not, by the major electronic camera manufacturers. Whether or not Arri's/RED/Sony's/Canon's LOG is a true log is a subject for another discussion and debate. But that doesn't matter. The idea is to retain as much dynamic range for later grading by using a LOG curve.

LOG has no relationship to actual RAW sensor data! True RAW capture does not store any data in linear or log - that's why it's called RAW. Sony's F3 S-LOG is just a gamma curve, as is RED Redfilmlog.

There's also a lot of misinformation in this thread about the relationship of bit depth to dynamic range, but that's another discussion.
 
yes I agree the idea of using a log curve is in order to preserve as much of the dynamic range capable by the sensor of the camera.

But early on REDCODE was disclosed by RED as being a linear based encoding file format, is RED not still recording linear data or is redfilmlog for example now using some type of gamma curve to it already?

cheers,
 
Last edited:
Yes, all video sensors that I know of are still by their design, linear. I meant a linear gamma in my previous post. These LOG gammas are intepretations of that raw "linear" sensor data. Redlogfilm is a logarithmic manipulation of that linear data. All of them are really, Log-c, log-s, canon log, since their sensors are by their nature, linear, the log curve in an interpretation so that the data can be available or rendered for later grading - usually baked into dpx or tif files.
 
bgundu said:
Just so the whole 20 stops doesn't get out of control, there is only 18 stops if you look at the chart. You ignore the first and last reading. Plus after calibration it will most likely be 16-17 stops on Dragon.

http://www.reduser.net/forum/showthread.php?92158-Enter-the-Dragon&p=1129432&viewfull=1#post1129432
Click to expand...

True Bob, good point.

I also did read somewhere in Jim's post that Dragon will still be REDCODE based, so that answers that. It looks like the sensor data will remain linear encoded, or in other terminology "linear gamma".
 
How is the ADC function performed on the or any other of the high end cameras ( eg the Sony F3 firmware upgrade for S log), if it's just a dedicated ADC chip then any S log function would be a software operation of the digital o/p or more likely the ADC function is performed by some sort of programmable PLA prog logic array type chip hence S log or whatever can be factored into the ADC conversion.
 
AFAIK, ADC is always linear, but you can then apply a curve before encoding, even if you're recording RAW.
On modern Sony stills sensors, this happens in the sensor itself: there's an ADC for every row (or column, I don't remember for sure). But each sensor can be designed in a different way.
 
I was under the impression that CMOS sensors generally were designed with amplifiers and ADCs on the chip. I'm sure there is some variance in this, though.

You guys are all right even if you don't all understand each other...

Generally in image sensors: photosite light is linear, charge is linear, voltage is linear, adc is linear, but then that digital signal can be encoded in many different ways, log, linear, gamma-encoded, compressed, uncompressed, etc. Of course there certainly variations in this as well, where even analog or adc conversions could be adjusted or made non-linear but it is not very common.
 
I don't honestly know how the imaging sensors are constructed, I m just speculating on how some manufacturers do it.
1- on the having the ADC's and amps on the same chip, shorter signal path less loose but increased noise and heat on the imaging chip.
2- any form of pre processing before ADC conversion will add noise hence the guess is the ADC function variable so that S log etc can be incorporated and updated via firmware tweaks.
As I have said I don't know, anybody care to shed some light/photons etc.
 
nyvz said:
I was under the impression that CMOS sensors generally were designed with amplifiers and ADCs on the chip. I'm sure there is some variance in this, though.

You guys are all right even if you don't all understand each other...

Generally in image sensors: photosite light is linear, charge is linear, voltage is linear, adc is linear, but then that digital signal can be encoded in many different ways, log, linear, gamma-encoded, compressed, uncompressed, etc. Of course there certainly variations in this as well, where even analog or adc conversions could be adjusted or made non-linear but it is not very common.
Click to expand...
Yes, this is exactly right, and what I've been trying to say all along, all hardware functions up until the digital encoding is linear. You see, most sources say not to compare bit-depth to dynamic range, because historically, the analog properties of sensors never even remotely reached the point where bit depth actually became the limiting factor... But this factor, for the first time in history, has been surpassed.

If you look at audio, it's the exact same deal, 16-bit allows for 96db of dynamic range, while 24-bit allows for 144db. In practice, people often argue the need of 144db, since the noise floor of even a super-high-end system in a studio space is around 90db. But sensors currently operate at significantly lower bit depths on average (10 to 14-bit), and bit depth does have the potential to be the limiting factor.

The way the Alexa gets it's clean 14 stops of DR is by using two sets of 14-bit ADCs that operate at different native gain levels (dual gain), and blends the result in-camera by taking the brighter range of the higher gain reading, and the shadow range of the lower gain reading. It's possible the Dragon sensor, assuming the DGA technology isn't patented, does something similar, but to a far greater degree, which is why the upgrade now needs a change of electronics as well. It could be thought of as HDRx that works by varying ISO, rather than shutter speed, but the requirement is to have ADCs that are natively calibrated to two different gain levels, which I assume is quite a challenge.
 
What nyvz said.

And also...

Kolor-Pikker said:
The way the Alexa gets it's clean 14 stops of DR is by using two sets of 14-bit ADCs that operate at different native gain levels (dual gain), and blends the result in-camera by taking the brighter range of the higher gain reading, and the shadow range of the lower gain reading. It's possible the Dragon sensor, assuming the DGA technology isn't patented, does something similar, but to a far greater degree, which is why the upgrade now needs a change of electronics as well. It could be thought of as HDRx that works by varying ISO, rather than shutter speed, but the requirement is to have ADCs that are natively calibrated to two different gain levels, which I assume is quite a challenge.
Click to expand...

IIRC, the sensor in the BMC does this as well. It was this one, right? (check Figure 3)
 
You must log in or register to reply here.
Back
Top