Mysterium-X dynamic range...

[Pawel Achtel]

It may be that simply downsampling the image from the existing sensor would give you similar noise characteristics to a purpose-built sensor of lower resolution but otherwise the same technology.

Put another way, if the sensor readout noise is low compared with the photon shot noise, then there would be no point to making a custom low-light sensor with bigger pixels- you'd get the same image by just downsampling in post.

Apparently not.

With low number of photons per pixel, averaging multiple pixels does not improve overall noise. Let's assume 100% QE. Imagine one photon falling onto each small pixel. The readout noise will be about the same as the signal level. Averaging 4 pixels, still results in the readout noise similar to signal level. Now, if we imagine pixels twice as large (4 times the area), each pixel would receive, on average, 4 photons, increasing S/N by about 2 stops. For the same reason, in low light, adding multiple exposures together does not work as well as taking single long exposure.

See some good examples here:

http://www.clarkvision.com/articles/night.and.low.light.photography/index.html

and here

http://www.clarkvision.com/articles/does.pixel.size.matter/

Indeed, I believe large pixels are the best way to reduce noise level in low light imaging. That's why we have cones and rods in our eyes and our vision is amazingly good, able to detect single photon in complete darkness as well as we are able to resolve amazing detail in full bright daylight. But, using different receptors (pixels) nicely overlaid together by our brains. :001_smile:

 

[Jannard]

There is a huge fallacy that only pixel size determines noise traits. If that were the case all 6 micron pixel sensor would perform roughly the same... they do not. In fact, there are pixel designs that improve noise performance with the same size pixels... Mysterium-X is a great example. Same pixel size as the original Mysterium sensor yet 2 stops less noise.

We have been told that our new sensor measures (noise) favorably to the new Arri sensor, which has much bigger pixels and much lower resolution.

And while pixel size is somewhat important, pixel design (and other things) is the new frontier...

Jim

 

[Deanan]

Aparently not.

With low number of photons per pixel, averaging multiple pixels does not improve overall noise. Let's assume 100% QE. Imagine one photon falling onto each small pixel. The readout noise will be about the same as the signal level. Averaging 4 pixels, still results in the readout noise similar to signal level. Now, if we imagine pixels twice as large (4 times the area), each pixel would receive, on average, 4 photons, increasing S/N by about 2 stops.

Transferring the charge out of a bigger pixel can also negate some of the advantages over a smaller pixel.

 

[Pawel Achtel]

Jim,

I assumed all other things being equal. Clearly, not all pixels are created equal. LOL If you can make significant improvements in quantum efficiency and pixel design, you cartainly may have a point. You can, indeed, achieve better result than large pixels of a different design and QE. No argument here, bring it on. :biggrin5:

What I was tring to point out is that, if pixel design and QE are the same, larger pixels will perform better (produce lower noise) at VERY low light levels. This is because we still have difficulty capturing fractions of photons in a single pixel. :smile5:

I think the easiest example to comprehend is adding several individual exposures vs. making a single long exposure. Same pixels. Single exposure will give a better result at very low light levels. Interestingly, at normal light levels, adding images will produce better result than a single exposure.

This is why I suggested two different pixel sizes. The larger, only for VERY low light levels (like single photons), which can be the case at very high speed cinematography or/and near complete darkness.

 

[Jannard]

Pawel... I understand your suggestion, although it is quite an expensive one. The point I would like to make is that no two pixel designs are alike... so you can't simplify quite so easily. What ultimately matters is the combination of many things, including pixel design, pixel size and a ton of other things that take you to the decision point. For us, it is a high resolution, high frame rate, low noise... record to Compact Flash solution. One that hasn't been matched in the industry. (Yet). Our goal is to keep it that way.

Jim

 

[Deanan]

Jim,

I assumed all other things being equal. Clearly, not all pixels are created equal. LOL If you can make significant improvements in quantum efficiency and pixel design, you cartainly may have a point. You can, indeed, achieve better result than large pixels of a different design and QE. No argument here, bring it on. :biggrin5:

What I was tring to point out is that, if pixel design and QE are the same, larger pixels will perform better (produce lower noise) at VERY low light levels. This is because we still have difficulty capturing fractions of photons in a single pixel. :smile5:

There's a big difference though between what theoretically might be better and realities of a hugely complex system from the pixel design to the A/Ds to the frame rates and other factors that impact the quality of the low light sensitivity.

Pixel size is too crude of a simplification to assume it'll always be better. From our experience building sensors and the enormous effort put into the getting the MX sensor where it is, pixel size alone is a small part of a very huge puzzle.

 

[Pawel Achtel]

There's a big difference though between what theoretically might be better and realities of a hugely complex system from the pixel design to the A/Ds to the frame rates and other factors that impact the quality of the low light sensitivity.

Pixel size is too crude of a simplification to assume it'll always be better. From our experience building sensors and the enormous effort put into the getting the MX sensor where it is, pixel size alone is a small part of a very huge puzzle.

Again, no argument here and the proof is in the pudding, or rather in the images coming from the MX sensor.

What about the temporal capability? Would a lower resolution sensor make it easier (or possible) to record to CF at, say 500 fps? Of course, there may be a better solution to that with a brain that could very quickly decimate the sensor output to either lower resolution or higher compression or both for recording.

 

[Jannard]

Again, no argument here and the proof is in the pudding, or rather in the images coming from MX sensor.

What about the temporal capability? Would a lower resolution sensor make it easier (or possible) to record to CF at, say 500 fps? Of course, there may be a better solution to that with a brain that could very quickly decimate the sensor output to either lower resolution or higher compression or both for recording.

The only time that the brain alone can increase frame rates is if you line-skip... which you know we think is a very bad thing.

Frame rates at full resolution (without line-skipping) are determined by sensor clock speed, read-reset times along with the ability to get the info off the sensor and through the system. If you take a smaller portion of the sensor... you can read it faster.

Jim

 

[Pawel Achtel]

Right. So, we should be able to achieve higher frame rates with the MX brain by simply reading out a window, just as we currently do with R1. Fair enough. Just need 16mm or 2/3" lenses but that's OK.

 

[Daniel Browning]

I hope nobody minds a few minor clarifications.

Aparently not.

With low number of photons per pixel,

FWIW, John Beale was specifically talking about tonal levels predominated by photon shot noise, which is not a low number of photons per pixel. Just FYI.

With low number of photons per pixel, averaging multiple pixels does not improve overall noise. [...] Averaging 4 pixels, still results in the readout noise similar to signal level.

No. The signal adds linearly, but the noise adds in quadrature (1*4/sqrt(1*4)), so the 1:1 SNR of each pixel becomes 2:1 in the resampled pixel, even at very low signal levels.

Now, if we imagine pixels twice as large (4 times the area), each pixel would receive, on average, 4 photons, increasing S/N by about 2 stops.

If read noise in electrons stayed the same in both pixel sizes, then the the SNR of the larger pixel would be double that of the resampled four pixels above (1 stop). But in practice, there are many cases where the larger pixel has correspondingly larger read noise. For example, the 5D2 got read noise down to 2.5 e-, then the 7D came out with much smaller pixels, and correspondingly smaller read noise (1.7 e- IIRC), so that the read noise per area is the same in both (i.e., noise power per spatial frequency in lp/mm).

This is because we still have difficulty capturing fractions of photons in a single pixel.

Actually, "fractions of a photon" are no problem. What that really means, of course, is that some pixels get a photon and others don't. There is nothing special about going below 1:1 SNR. I've resampled images that were 1:8 SNR (1 photon for every 8 noise electrons) down to web-sized images and you still get usable signal. For example, here is a 5D2 image at ISO 1 million, resized from 21 MP to 0.2 MP:

The extreme reduction in size made it possible to still get detail. If you looked at the 100% crop, it looks like just noise. On average, each pixel is only getting a "fraction" of a photon.

I think the recent M-X video was good indication that whatever technology RED developed is going pretty well.

 

[Pawel Achtel]

Actually, "fractions of a photon" are no problem. What that really means, of course, is that some pixels get a photon and others don't. There is nothing special about going below 1:1 SNR. I've resampled images that were 1:8 SNR (1 photon for every 8 noise electrons) down to web-sized images and you still get usable signal. For example, here is a 5D2 image at ISO 1 million, resized from 21 MP to 0.2 MP:

...

The extreme reduction in size made it possible to still get detail. If you looked at the 100% crop, it looks like just noise. On average, each pixel is only getting a "fraction" of a photon.

Thanks Daniel, your input is always enlighting.

I didn't say we could not capture "fractions of photons". We can because of the statistical distribution over larger number of pixels and your image shows that. But, it is not as good as capturing with larger and fewer pixels (all other things being equal). Averaging low photon distribution is not as good as higher photon distribution. And, a good comparison of temporal averaging vs. higher photon count per pixel (here over longer time rather than over larger area) demonstrated by Roger Clark here:

http://www.clarkvision.com/articles/night.and.low.light.photography/index.html

For this reason, I thought, larger pixels would work better with high speed and very low light. But, obviously, this is when all other things are equal. Clearly, MX sensor pixels are not equal to any other pixels I'm used to :smile5:

 

[Rob Ruffo]

There is a huge fallacy that only pixel size determines noise traits. If that were the case all 6 micron pixel sensor would perform roughly the same... they do not. In fact, there are pixel designs that improve noise performance with the same size pixels... Mysterium-X is a great example. Same pixel size as the original Mysterium sensor yet 2 stops less noise.

We have been told that our new sensor measures (noise) favorably to the new Arri sensor, which has much bigger pixels and much lower resolution.

And while pixel size is somewhat important, pixel design (and other things) is the new frontier...

Jim

For proof of this we need only look back to the first models of DSLRs produced in the early 2000s. At 4-5 megapixel (full-frame still 35mm - so HUGE pixels) they exhibited pretty much unusable levels of noise at ISOs higher than 300-400.

 

[Rob Ruffo]

Wow Todd, that was pretty harsh, and incorrect on your part. And depending on which Macbook Pro you are looking at, you have either a 6bits/channel or 8bits/channel display. If it's the LED backlit model, it's 8bit.

Graeme pretty much sums it up. We're finally starting to see affordable 10bit LCD panels trickle onto the market now. The HP Dreamcolor is somewhat 10bit, there are a handful of more expensive grading and coloring panels out there. A few of the new high-dollar LED LCD HDTV panels are shifting to 10bit. LG's 47LH90 HDTV is reportedly 10bit if I've read their spec-sheet correctly and it sure looks nice, rivaling a nice plasma in most respects. But it's expensive for a < 50" HDTV, although a real bargain for 10bit, if it truly is (I haven't confirmed).

My problem with all LCD tech (inlcuding LED - because it's just an LCD with LED backlighting) is refresh speed (still adds blur that is not really there, less than before, but some, and nowhere near the specs they lie to you with) and viweing angles. CRT and plasma do not vary color at all based on viewing angle, and this is super-impportant for final grading. For this reason all our worstations have a second CRT monitor for reference.

 

[Rob Ruffo]

Static contrast ratios with the ANSI checkerboard pattern are the only ones worth talking about IMHO.

Graeme

Agreed.

 

[jbeale]

theory vs. reality

theory vs. reality

http://www.clarkvision.com/articles/night.and.low.light.photography/index.html
http://www.clarkvision.com/articles/does.pixel.size.matter/

Thanks for those links- those are very informative pages! They show the 1Dmk2 has significant readout noise, so spreading the same photons among more pixels degrades overall SNR. So my theoretical argument assuming readout noise low compared with shot noise (R.N.Clark distinguishes readout + electronic + A/D noise) is shown to be false for 2004 sensor technology (1Dmk2).

I understand that readout noise is getting lower as the years pass and technology improves. Photon shot noise is set by physics but I think readout noise is just* limited by technology. The better it gets, the less advantage a large pixel has over bunch of smaller ones of equal total area, summed together.

I gather we haven't reached the negligible-readout-noise regime yet. But we may be, gradually, headed that way.

*A Mere Matter of Engineering(tm) as the physicists remark, annoyingly to the engineers.

 

[joshua csehak]

As for bit rates, that is a measure of bandwidth or transmission speed.

Haha I meant bit depth. I should say what I mean

But thanks for all the great info. I have a feeling sensor design is one of those things where the more you understand, the less you understand.

 

[Xander Fischer]

Leica f/.95 and REDmx... Yummy, I bet.

Leica f/.95 and REDmx... Yummy, I bet.

13 stops of DR is absolutely fantastic.

Looking forward to test R-1-M-X soon with Leica Noctilux-M 50mm f/0.95 at wide open @ f/0.95.

:shocked:

PLEASE send me a link to that. I have yet to see Leica glass on RED... much less and video camera... Let alone a REDmx.


That'll be interesting.

 

[Eki Halkka]

The only time that the brain alone can increase frame rates is if you line-skip... which you know we think is a very bad thing.

It would be cool though, if it would be possible to do an interlaced read of the sensor, for double frame rate. Of course, saving interlaced scan as raw probably wouldn't work too well...

Anyway, in the odd case it's possible to scan and save interlaced, there's some pretty decent post production software to de-interlace footage with, only in moving areas of the image.

The result is lower quality than true progressive, but at times the quality loss could be acceptable tradeoff for higher frame rates.