Red Compression and Matrix Tests.

[Graeme Nattress]

There will be point where the number of photons entering the pixel is not great enough to make a signal larger than the noise. This would be the case with any sensor. That's why bigger pixels are better in low light, as they let more photons in. They also have a larger photon capacity, which leads to larger max signal, and hence better SNR and dynamic range.

Graeme

 

[Brook Willard]

This is the best thread ever...

 

[Andrew M.]

1 stop = 1 bit = 6db, roughly speaking.

Graeme, what is the definition of the stop?
I know stop on the aperture, but in context of this thread?
Somehow I can't get this expression of sensitivity in terms of stops.
dB Yes
bits Yes
dynamic range no problem
Stops ?? can't get it.

Andrew

 

[Andrew M.]

So the sensor has a 2500 step linear resolution, and the dynamic range of the chip goes up to a little over 2000 times brighter than the darkest object it can recognize. Is the noise behavior of the chip a constant, or is there less chip noise in the low-end or high-end?

Good sensor will have similar characteristics as the human eye.
As all in the nature we use, is logarithmic. Intensities, loudness, brightness, concentration of chemicals in our body, all is logarithmic.
So it has much finer resolution at low levels and less resolution (in terms of steps or increments) at higher levels.
Gamma curves try to deal with this if the system do not have it implemented right in the pickup (sensor) or in the display (screen).

So if the sensor do not behave logaritmic then analog to digital converter better have finer resolution at low levels then at high levels. It is a specially true for audio.
With RGB or light levels it is bit more complicated but our eyes do have less sensitivity response at high levels.
The question is, as a industry, what is the standard, are we dealing with the non linear nature of human eye only on the screen device or on the sensor as well.
My intuition is that we should use 80% of the bits resolution on the first half of the signal and the 20% on the second half.
Andrew

 

[Mark B.]

Graeme, what is the definition of the stop?
I know stop on the aperture, but in context of this thread?
Somehow I can't get this expression of sensitivity in terms of stops.
dB Yes
bits Yes
dynamic range no problem
Stops ?? can't get it.

Andrew

The way I understand it, a stop is twice the amount of light as the previous amount (the previous stop). If you're talking about sensor sensitivity to stops, you're talking about a change in the amount of light hitting the sensor. If a sensor captures 2 stops, it would be able to capture everything from it's lowest sensitivity up to 4 times that amount of light. Picking a number at random for an example... if the lowest light captured is 6 lux, then the 2-stop sensor would be able to detect light levels all the way from 6 lux up to 24 lux (6x2x2). If that sensor had 3-stops of range, it could detect up to 48 lux (6x2x2x2). 4-stops and it could detect from 6 lux up to 86 lux (6x2x2x2x2). Anything above the 86 lux would be blown highlights. If the sensor has a different minimum sensitivity, say 2 lux instead of 6 lux, then the 4-stop range would go to being 2 to 64 lux (4x2x2x2x2).

But maybe my math is wrong... it's late.

 

[Mark B.]

Good sensor will have similar characteristics as the human eye.
As all in the nature we use, is logarithmic. Intensities, loudness, brightness, concentration of chemicals in our body, all is logarithmic.
So it has much finer resolution at low levels and less resolution (in terms of steps or increments) at higher levels.
Gamma curves try to deal with this if the system do not have it implemented right in the pickup (sensor) or in the display (screen).

So if the sensor do not behave logaritmic then analog to digital converter better have finer resolution at low levels then at high levels. It is a specially true for audio.
With RGB or light levels it is bit more complicated but our eyes do have less sensitivity response at high levels.
The question is, as a industry, what is the standard, are we dealing with the non linear nature of human eye only on the screen device or on the sensor as well.
My intuition is that we should use 80% of the bits resolution on the first half of the signal and the 20% on the second half.
Andrew

If a sensor captures light with constant resolution, regardless of the light intensities falling on it, then it doesn't make sense to collect data in a logarithmic way. That would only result in a loss of perfectly good data on one end and the collection of redundant information on the other end. If it's a logarithmic sensitivity, then it makes sense to set the data collection to a logarithmic weight.

BTW, what chemicals in our bodies are logarithmically concentrated?

 

[david farland]

I thought most of it was....at least in our sensory perception!

It only matters if one part of your system is logarithmic i.e voltage output of the photo diodes collecting photons....but hey even it it wasn't you could change the tranfer characteristics in software to closer match visual perception.

DF

 

[Policar]

Just thinking:

Film's characteristic curve has about 8-10 stops (or whatever) of linear information, and a smooth roll-off at the highlights (and, with some stocks, also in the shadows.) I bet Red has more linear dynamic range than comperably fast (ISO 500, exposed as ISO 320) film stocks, it just lacks the smooth knee. (Personally I'm not a huge fan of shadow detail; I love the look of Michael Bay movies and (so far as film stocks go) nature photography using Velvia.)

Both the shadows and the highlights are so bright or dark as to have little visual information in them, and the extra dynamic range of film in these areas cannot effectively be recapture in the lab, because they are not linear. But they still look good!

Would it be possible to make a codec that's primarily linear, with logarithmic ends at the knee and in the shadows (where the S:N ratio is so low it wouldn't matter anyway)? Sure, the signal would be poor in shadows and highlights, but so too is it unusably poor in film.

 

[tj williams]

Policar

Possibly gamma curve tools in RED Cine will address this at the output stage?

 

[Rob Lohman]

You can pick an output gamma space when encoding in REDCINE

 

[P Andersson]

This is the best thread ever...

absolutely, it was interesting to study how film reacted to light a few years ago, (remember the zone system?) and it is essential to build a strong foundation in understanding the stuff of this thread, this information will sit in the bone marrow eventually and determine automatic adjustments to the f stop in critical situations

You can pick an output gamma space when encoding in REDCINE

do you guys have some rough samples of how it looks when you have encoded to different gamma curves that you can post for us to learn the difference, we could get a visual feel for how noise reacts in different situations and what gets clipped when etc.

if not, could I do this with any raw file in photoshop somehow

 

[Graeme Nattress]

The eyes response to light is not logarithmic though - it's actually a close approximation for the power law curve of a CRT ~0.4.

Imaging sensors are, however, inherently linear and noise limited. Also, mathematics like colour transforms must be applied to linear, not gamma encoded data. So, if you sampled the data logarithmically, the first thing you'd need to do would be convert it back to linear to do your colour science math on it.

Power law and logarithmic curves can make sense for compression, however. However, with RAW you've got to be careful as you don't have an implied rendering intent with regards to how you want to deal with highlight and shadow data, so it's wise not to push the curve compression too far. All you've got to remember, is that image processing works best on linear data, they eye has as power curve of about 0.4, and that there's a lot of stuff that goes on inbetween....

Graeme

 

[Andrew M.]

If a sensor captures light with constant resolution, regardless of the light intensities falling on it, then it doesn't make sense to collect data in a logarithmic way. That would only result in a loss of perfectly good data on one end and the collection of redundant information on the other end. If it's a logarithmic sensitivity, then it makes sense to set the data collection to a logarithmic weight.

BTW, what chemicals in our bodies are logarithmically concentrated?

Lowkus, thanks for the stop explanation, finally I got it.

Correct me if I am wrong but most sensors will have linear bottom curve with saturation showing up at the top at the maximum illumination.
Analog to Digital converters could be modeled to convert analog sensor output to maximize its bit significant resolution. Now 12 bits digital output from the sensor has 4096 steps. The steps should be distributed in such a way so to reflect significant changes of the brightness and/or chroma for each step change, we don't want to waste bit steps for nothing, do we?.
Where significant mean significant for the human eye. After all we are producing output to be looked at.

So we could apply gamma curve right at the sensor A/D converter or to build the photo sensor that is modeled on the human eye sensitivity curve, or both.
IMHO closer we are to the real thing (the eye) then better.

I remember all these sound quality issues 20 years ago, finally at 16 bit sound resolution no body argue anymore that electronic tube based amplifier sounds better then digital one
Microphones are build to respond like human ear and headphones are trying to pass this signal without any distortion. I will not touch the speakers issues since this technology is far far behind the minimum requirements.

I hope that in near future the camera sensors will consist of two eye like devices with the same characteristic.

BTW the logarithmic concentrated chemicals in our body:
blood sugar levels, hormones, etc.
Small change of blood sugar level at low mark will make us feel better, doubling it again and again will not produce the same reaction. Kind of like sex, logarithmic, hmmm..

Andrew

 

[Andrew M.]

The eyes response to light is not logarithmic though - it's actually a close approximation for the power law curve of a CRT ~0.4.
Graeme

Graeme, you are perfectly correct.
By saying logarithmic, I mean power curves.
Kind of like using dB to express levels.
3dB increase of power it mean double of power.
6dB increase of power level is actually 4 times of power level.
9dB increase of power level is actually 8 times of power level.
Bit different in the levels/voltage
6dB increase of level (voltage) it mean double of voltage.
So here we have 6dB to double.

Is this logarithmic or power play?

 

[Graeme Nattress]

Andew, sensors are inherently linear - that's how they work. 1 photon = 1 electron charge = analogue voltage = digital number.

Now, if 12bits gets you down below the noise floor, then you're already below the smallest needed step size. You can't now distribute those 4096 steps to give more detail to the shadows and less to the highlights as you're already with the finest, smallest step size that makes sense for the noise floor of the sensor. You could distrubute LESS bits so that you get 12bit resolution in the shadows and bigger step sizes in the highlights, though, but the easiest way to do that is to start with linear data and add a tone curve. Much easier than trying to build a non-linear a-to-d.

I have no idea what you're going on about tube amps and digital amps and 16bit and all that. 16bit implies a 96db SNR which both tube and transistor amps can better, and quite frankly, you couldn't pay me to listen to music on a transistor amp

As I pointed out, the way the brain percieves brightness is on a ~0.4 power curve, not a logarithmic curve.

Microphones are designed to be flat (ie linear) and not with built in Fletcher–Munson curves. Nobody puts "loudness" countours into amps any more either

Graeme

 

[Graeme Nattress]

Logs and power law curves are related, but different

 

[Andrew M.]

Andew, sensors are inherently linear - that's how they work. 1 photon = 1 electron charge = analogue voltage = digital number.

I have no idea what you're going on about tube amps and digital amps and 16bit and all that. 16bit implies a 96db SNR which both tube and transistor amps can better, and quite frankly, you couldn't pay me to listen to music on a transistor amp

Graeme

I didn't know that CMOS sensors are linear. I guess I have much to learn.
I am out of audio business since 15 years, but I remember how serious audiophiles did buy tube based amplifiers than (still you can buy one) saying that they do sound much better.
Better, maybe YES for some, accurate reproduction, NO.
I am believer in math and in digital world so S/N of 16 bit is probably the max what human ear can detect but if this will not be enough at current state of audio technology we can easily triple or quadruple that.

 

[Graeme Nattress]

Yes, can still buy tube amps. I've got a couple. They sound different. It's a film v digital cinema thing, where it's a user-choice aesthetic. Tube amps are expensive, large and hot, but they do sound rather nice, if not accurate. Of course, I'm a vinyl head too - same argument - they sounds nice, if not accurate.

16bit is quite adequate for playback, given the usual 50db or so noise floor in a domestic environment and the 96db dynamic range of 16 bit. To hear it all, you've got to be in an anechoic chamber, or rapidly become death, or both. However, I think >16bit is very useful for recording / mastering though to give you some "headroom".

CCD and CMOS are both very linear. It's a direct transform of photon to electron to voltage. Some non-linearity would be very useful for capturing highlights, but that's still an active area of research.

Graeme

 

[Andrew M.]

Graeme, I just realized something, are you serious audiophile?
If yes, my apology, I did step on the religious and very delicate subject, also this discussion is not only for a different thread but for a different forum all together
Sorry.............
Andrew

 

[Graeme Nattress]

Yes, quite serious But also very balanced. Check out my Stereo Images movie here: http://www.nattress.com/Movies/movies.htm

Audio can be a religious topic, but I'm a mathematician, so I fully understand why a CD player is more accurate than vinyl, how sampling theory really works etc. However, I recognise that a lot of the music I like was mastered with vinyl playback in mind, and hence, although less accurate to the master tape, it's probably more like what it was intended to be heard like. Vinyl is also more "fun", but as I say, less accurate. Similarly with tube amps, they're fun, and they have a life to the sound that is wonderful to listen to, but they're less accurate than a transistor amp with a bunch of feedback. I do think this would belong in a off-topic "philosophy" thread on accuracy v enjoyment though.

Graeme