16-bit linear vs 10-bit log?

[Rocket]

Hi Guys,

Just a question. I was reading through the DALSA specs and was quite surprised by the 16-bit linear RAW output from it's sensor. What real-world difference does this make in comparison to the 10-bit log we will get from the Red?

What was the reasoning behind not going to 16-bit with the Mysterium? Does it have to do with the fact the DALSA sensor is a CCD and the Mysterium is CMOS?

Rich

[Bruce Allen]

You can also record 12-bit linear from the Red. It's not a huge difference to 10-bit log - because the log curve puts the bits in the right places.

I think the rationale behind Red's 12-bit recording would be that when you get to a certain point, superior bit depth is not necessary because all you'll be recording is noise... there's a law of diminishing returns there, versus putting resources into another part of the sensor / data pipeline. It's very similar to audio recording - 8 bit vs 16 bit is a huge difference, 16 bit vs 24 is noticeable but not as huge. People don't go beyond 24 for now because the money is better invested in mics / preamps / etc...

Of course, the Dalsa sensor is very clean and from the tests I've seen them show, the images are really great. Maybe their sensor is a little bit cleaner than the Red's and they felt it needed a couple more bits to be absolutely sure of storing everything it captured?

On the other hand, the Red stuff that Jim's been posting recently seems to be able to be pushed quite a few stops without quantization. Of course, we need to see full-frame moving images to tell...

Bruce Allen
www.boacinema.com

[Graeme Nattress]

I've not seen noise specs from Dalsa that would indicate that 16bits is necessary to record it's dynamic range. The good thing of recording lots of bits is you're certain to get all the range of the sensor, but the bad thing is that if you record too many bits you waste bandwidth and disk space. If you "nail" the number correctly, you get both the full range of the sensor and you don't waste space or bandwidth.

Graeme

[J. Bernard Vallon]

I've noticed that all the cameras i've seen with dalsa sensors have 16 bit outputs. The digital hasselblad backs, the origin, and that huge seitz 6x17 digital panoramic camera. It made me wonder if thats just how dalsa likes to roll, or if its really necessary.

[combatentropy]

10 bits per pixel is like having 1.07 billion crayons to color your picture with: (Two to the tenth power, for green) times (two to the tenth, for blue) times (two to the tenth, for red).

16 bits is 281.5 trillion. That's a lot more, but the human eye cannot see the difference. Plus, you could only hold 11 minutes on a 320 GB drive.

Plus, when you put your movie on DVD, it doesn't matter, because the DVD guy is going to copy your picture with his set of only 16.8 million crayons (8-bit).

The advantage of more bits in your original is for stretching the picture in post. For example, if you brighten your picture something crazy like to ISO 4000, like Red has done in another forum post, it decreases the overall palette. In this case, I think, from 12 bit to 8 bit.

This is because you are basically saying, "Okay, I am going to redraw the original, but for each color, I am going to instead switch to the crayon four shades brighter. So, for orange in the original, you redraw with your light, light, light, light orange. For dark orange, you switch to light, light, light orange. For dark, dark orange -- light, light orange. And so on. . . . The turning point comes when you get to four shades below your white crayon. It is replaced with white. So is the color three steps below white, the color two steps, the color one step, and white itself. They all become white. So, a decrease in the total breadth of your palette.

Eight bits is still acceptable for distribution, but at ISO 4000 and above, other limits creep in, like the noise floor of the sensor. My point is, unless you regularly need to push your picture six or seven F-stops, and all of the other advantages of Red over Dalsa don't matter (portability, modularity, cost), then maybe Dalsa is the one for you. Otherwise, it would be a waste.

[J. Bernard Vallon]

combatentropy,
good analogy, i'll probably reference it to explain color space transfers in the future. You do gain a benefit by having a high bit-depth besides pushing and brightening. If your going for a complicated look, like in saving private ryan or something like 300, where you are actually only looking at a few small sections of the original color space, your resulting image will look much cleaner with less pasteurization if you master in higher bit-depth.

[combatentropy]

less pasteurization if you master in higher bit-depth.

Less "pasteurization," eh? :w00t: Good, I hate that milky look.

[Rocket]

Thanks guys, that helps. So I take it there actually is going to be no noticeable difference in the end if the scene you are shooting is lit correctly for the exposure (no need to under or over expose at all). All the image information you need (and more) will be captured to hard disk even at 10 or 12-bit.

The explanation of pixel size is also helpful, especially when it comes to understanding photodiode/lenslet and fill factor.

[combatentropy]

So I take it there actually is going to be no noticeable difference

This rectangle is actually two squares beside each other. The right square is one shade away from the left, using the next step available in the 8-bit color space.

http://www.reduser.net/forum/uploaded/2100_1184695994.png

Can you tell whether it is one step lighter or one step darker?

With 10-bit, you will have six more shades between these two.

With 12-bit, you will have 14 more.

Eight bit is about the threshold for the human eye, which I think is why distribution formats like DVD and your computer monitor have adopted it. More is invisible.

[betty schaefer]

i see banding all the time on dvds..mostly in scenes with a subtle dégradée.
try the same in greys and i think you'll see it

[Adrian T.]

Can you tell whether it is one step lighter or one step darker?

It's one step bluer. :sarcasm:

[combatentropy]

i see banding all the time on dvds

Sometimes you will see blotchy color in DVDs, digital cable, and photographs on a computer. That is due to compressing the image with 30-to-1 ratio with MPEG-2 or JPEG. Uncompressed 8-bit does not have such banding.

Redcode is compressed, but it is less (only about 12-to-one) and of a newer, better kind (wavelet instead discrete cosine transfer).


It's one step bluer. :sarcasm:
You cheated.

In this example, the right square is 3 steps brighter (using only the 8-bit color scale) than the square on the left.

http://www.reduser.net/forum/uploaded/2100_1184698126.png

10-bit would have 12 shades, instead of three, between these two squares.

12-bit would have 48.

[betty schaefer]

a section of a png 8 bit

http://img252.imageshack.us/img252/4716/snap247cb2.png

what's a darker red? my skirt or my blouse? :D

[J. Bernard Vallon]

Less "pasteurization," eh? :w00t: Good, I hate that milky look.

Damn spell check. Thats what I get for typing too fast.

[GlennChan]

Eight bit is about the threshold for the human eye, which I think is why distribution formats like DVD and your computer monitor have adopted it. More is invisible.
8-bit (with video gamma, normal gamut) is good enough in (the vast?) majority of situations. This is likely evidenced by the fact that LCD manufacturers almost get away with 6-bit panels.

In lab situations, we can see better than 8-bit.


>At the same time, 8 bits per channel is well below the human threshold >of color-difference perception

While this may be a commonly held belief, it does not hold up to rigorous testing.

In October 03, we (under request from DCI) did a comprehensive discrimination threshold test using theatrically projected images - at 40, 4, and 0.4 cd/m^2 using a gamma of 2.6. These were projected images on a 45' screen at the DCL in Hollywood. We tested approximately 75 subjects, from both "expert viewers" and non experts. The test patterns created steps in luminance corresponding to 1 count in 8,9,10,11 and 12 bits. The viewer was asked to identify the orientation of the test pattern, to verify if he really saw the steps.

The results of the experiment will be published in the September 04 SMPTE journal, but can be summarized by the following :

* Gamma 2.6 best follows the threshold of human perceptual discrimination
for luminance at cinema levels
* All viewers could see 8 bits all the time
* Almost all observers could see steps at 10 bits
* Some observers (40%) could see steps at 11 bits
* Essentially no observers (1%) could see steps at 12 bits.
* Observers learned to see steps and became more sensitive with repeated viewing

These differences calculated to very small DeltaE* - in the smallest case, the threshold of discrimination was 0.05 DE* (if I remember correctly), much smaller than previously thought to be visible.

This was the experiment that caused DCI to adopt a 12 bit (gamma 2.6) recommendation.

If the display does not exhibit contouring at 8 bits, this will be caused by several factors :

* Noise - (either in source or display) dithering will hide contours very effectively

* Environment - environmental contamination of the display (e.g. lights on around display)

Note that if you translate this into linear space, it requires a lot more bits.

Matt Cowan
http://www.cinematography.net/Pages%20DW/ColorBitDepth.htm

2- In the grey example above, you can see the difference if you zoomed in or if the patches were bigger. The noise in that image helps obscure things though. Visual angle makes a difference... sensitivity goes up for larger angles.

3- There's a difference between 12-bit linear and 12-bit gamma of 2.6 in DCI. (12 bits linear does not give a "perfect" 12-bit gamma/2.6 signal; Small amounts of noise might smooth things out to give visually perfect results perhaps.) Though in practice I wouldn't worry about it.

[Blair S. Paulsen]

Good thread. The Mysterium sensor (as described) is a significant achievement but it is not the "perfect imager" or the holy grail. Some future sensor, probably combined with a 16 bit pipeline will smash all the performance benchmarks of the RedOne system (as described). I even buy the premise that the Dalsa Origin can outperform the RedOne in meaningful areas of image quality.

Of course back here in the real world the image quality that all tests to date indicate we can expect from the RedOne is legitimately "state of the art" - in a compact package, with a post pipeline that runs on commodity hardware and sensitivity/signal to noise capabilities that match or exceed any other motion imaging option I am aware of at any price.

Did I mention the plethora of operational and workflow advantages the RedOne has? That I agree wholeheartedly with Graeme's post about the 12 bit linear / 10 bit log choice as just right for sane bandwidth management without significant impact on the finished image - even for theatrical. How about the best implementation of wavelet compression I have ever seen, by a noticeable margin.

Oh yeah, I almost forgot - it hardly matters what kind of images you can acquire with a camera if you can not realistically acquire the camera in question. The day the RedOne ships (assuming they hit their schedule) it will change the price to performance benchmark of the industry by a full order of magnitude.

[Häakon]

10 bits per pixel is like having 1.07 billion crayons to color your picture with: (Two to the tenth power, for green) times (two to the tenth, for blue) times (two to the tenth, for red).

16 bits is 281.5 trillion. That's a lot more, but the human eye cannot see the difference. Plus, you could only hold 11 minutes on a 320 GB drive.

Plus, when you put your movie on DVD, it doesn't matter, because the DVD guy is going to copy your picture with his set of only 16.8 million crayons (8-bit).
The differences you pointed out between the bit-depth is accurate and helpful for comparison when everything else is equal, however the topic was querying linear vs. log in addition to the bit-depth. In this case, the numbers are vastly different.

[Michel Hafner]

Sometimes you will see blotchy color in DVDs, digital cable, and photographs on a computer. That is due to compressing the image with 30-to-1 ratio with MPEG-2 or JPEG. Uncompressed 8-bit does not have such banding.

Unless you apply dither 8 bit can very well have banding with critical gradients, uncompressed or compressed.

[Jannard]

This is a perfect example of getting lost in the specs. What does the image look like? Dalsa makes a great image but I don't recall any features being shot on that camera... not sure why. I do know that Peter Jackson compared the RED image to 65mm film, and that is with our 5.4 micron pixels. That seems good enough to me for a $17,500 camera that you can actually own. We have 4 features in the works and these will all be shot with prototypes. Getting into 16bit vs. 12 bit debate seems like an interesting discussion, but it just might lead to the wrong conclusion.

Jim

[Poi Boy]

Right on Jim ! You can't make it any more clear than that. I suspect Red is going to make the current generation of Dalsa into a very short term player.
Aloha
-A

[Gavin Greenwalt]

I'm extremely doubtful anybody is capturing a half-float's worth of dynamic range or color accuracy. Dalsa is probably just shooting to 16bits because there are no official spec'ed formats to run 12bit linear. You have DPX and Cineon for your 10bit Log work and let's not forget that's been used for all 13+ stops worth of data on film for the last 10+ years.

We won't need 16bit images coming out of the camera until we're capturing HDR imagery.

[Poi Boy]

Dalsa will need to adjust to a Red centric market very quickly or they will become obsolete.
Aloha
-A

[Rocket]

That seems good enough to me for a $17,500 camera that you can actually own. We have 4 features in the works and these will all be shot with prototypes. Getting into 16bit vs. 12 bit debate seems like an interesting discussion, but it just might lead to the wrong conclusion.

I agree entirely, but the interesting discussion aspect was half the point, it's the technology behind it that I am interested in at this level, and the reasons for or against doing something a particular way on a design level to achieve a particular result. If you take into account the majority of the responses to the question, I don't think anyone here is naive enough to look purely at the numbers and conclude that capturing 16 bits of information must be better than 10 or 12 bit log just because 16 is a greater number.

Just as was pointed out in sampling audio, there is a threshold at which the law of diminishing returns takes over and it is not necessary to spend time and money capturing information that takes up unnecessary bandwidth and storage and will never be used.

I think the majority of responses have been very objective and balanced in taking into account the fact that there is more to it than just pure bit depth.

I don't think anyone is going to post a "put the camera on hold! please re engineer the Mysterium to give us 16-bit sampling, wait, no why don't you supersize me to 24-bit" thread, "oh, and we want true lossless compression and an option to record a full 60 mins on a USB memory stick, thanks!"

Jim, you have an awesome community of people here, I am blown away by the level of technical understanding and openness in those who post that are not engineers (I am sure some are), and the honest, quick and helpful response of your team.

The number of people I've come across who can't even set the white balance on a PD150 (and have no concept of colour temperature anyway), but consider themselves professional camera operators is rediculous.

Sure, perhaps there are some that have ordered a Red and will be coming to it from Sony Handycam world, but it seems to me that the majority here will be combining a state of the art camera with far above average knowledge (and understanding... two different things) of the technology itself, of the entire post process, and the creative elements of lighting, framing a shot, exposure, focus, and actually using the camera to capture top class images.

So I don't think there is any reason for concern.

[Graeme Nattress]

Just because you capture 16bit, does not mean that you capture 16bit of information! Say you digitized VHS at a 16bit depth rather than 8bit depth. You don't get 8 bits of information extra, but just 8bits of noise.

What you need to do is get a Dalsa and measure the information content of the 16bits. You can do this by measuring dynamic range and what a test chart 7 stops under-exposed looks like. They'll tell you what you need to know.

Finally, and no less important, is not just a measure of the noise, but a visual look at it to see what characteristics it has. I'd encourage everyone to download the test images from Dalsa and look at them critically, especially in Photoshop where you can zoom in and also look at red, green and blue seperately. That's what we do when we look at RED images in development, and it's enlightening. Also, go over to CML and download their DPX files for the Genesis, F23, Viper and D20. You'll see "amazing" things in those images......

In the end, it's all about overally image quality, and the form factor that allows you to get the shot.

Graeme

[Bill Anderson]

These "other" camera names are tossed about with such gay abandon you'd think they were a practical alternative to RED. They are not- never will be. They'll be nothing but a two bit memory once RED hits the streets. The only thing RED owners have to worry about is being trampled to death by the latecomers. I can almost hear the sound of these "other" cameras hitting the ground, like so many hot potatoes. Last century it was Kodak, this is RED'S era.

[Graeme Nattress]

Bill, nice of you to say so, but we should never be complacent. People have made great images with "others", and from looking at what they did right and wrong we can learn.

Graeme

[Bill Anderson]

Graeme, I wasn't saying it to be nice, I wasn't even saying because I wish it, I was saying it because I believe it- and this from a guy who has Hydroquinone for blood!
Complacency? Heck no, but I don't believe RED has contenders to date, It's what's coming up the pipe that's going to define RED. So, by all means compare cameras, but when comparing please consider (apart from a few extra bits) the practical things like, cost, accessibility, sensor size, and whether or not the damn thing is tethered to a filing cabinet with a team of surgeons operating it. Dalsa? Viper? Gimme a break, guys. RED is an Indie camera that is about to make everything else indie- or obsolete.
Keep spurring.

[Roxco]

I agree entirely, but the interesting discussion aspect was half the point, it's the technology behind it that I am interested in at this level, and the reasons for or against doing something a particular way on a design level to achieve a particular result...

Sure, perhaps there are some that have ordered a Red and will be coming to it from Sony Handycam world, but it seems to me that the majority here will be combining a state of the art camera with far above average knowledge (and understanding... two different things) of the technology itself, of the entire post process, and the creative elements of lighting, framing a shot, exposure, focus, and actually using the camera to capture top class images.

So I don't think there is any reason for concern.

Just as DV and FireWire saved many a college/school's video gear budget, RED and FCP should save many a Film School's gear budget and may even let a few private high school's dabble in high rez productions.

Access to cool tools excites the imagination and seeing what a student can do with a DVX-100 or a RED One should always be a great joy for the Red Nation.

IMHO, the difference between knowledge and understanding is experience.

Rosco

[Steve Freebairn]

This has been a really interesting thread. I'm glad Red chose what they did and is making the camera more available(affordable). The images (no matter what the micron size of the pixels) are incredible at 4k. There have been complaints out there about 4k being too good and that people prefer 2k projection to 4k. Sounds like resolution isn't going to be a factor. The range is great, obviously in Film you can pull more data out of overexposed shots, but I think Red is definitely there with underexposed material. There is no sense in capturing too much data since we have to store and backup everything we shoot!

By the way NanoHV, did you see Crossing the Line projected at 4k? or just the super compressed downloads that were available? Even apples 2k version paled significantly in comparison to what Jim was showing in the booth.

[Jarred Land]

For what its worth, the Red Prototype "Franky" recorded to 16 bit frames back in the day. We learned alot.

We are now, with less bits, recording much better images.

And thank god, as Graeme mentioned, I was getting a little pissed off buying new terabytes every day just to record noise.

[Graeme Nattress]

Yes, Frankie used to spit out 16bit tiffs, but 4 bits were "blank" so to speak, leaving 12bit data. Wasted tonnes of hard disk space.

Now we're recording to compact flash and just laughing about the old refrigerator days....

Graeme

[Rocket]

RED and FCP should save many a Film School's gear budget and may even let a few private high school's dabble in high rez productions.

Access to cool tools excites the imagination and seeing what a student can do with a DVX-100 or a RED One should always be a great joy for the Red Nation.

IMHO, the difference between knowledge and understanding is experience.

Rosco

Totally, experience goes a long way, and it won't be long before a lot of people get to jump into production with this new technology and gain as much of it as they can.

I'm not sure I am quite with you on the Red + FCP allowing dabbling in high res production. An uncompressed conform even at 2K is going to require a lot more than FCP and a Macbook Pro, not to mention a Bluefish 4:4:4 2K card, and 4K forget about it. FCP is fine as an offline edit tool with SD or HD compressed proxies of your Red RAW 4K media but it's not going to help when it comes to handling your online conform, grading and finishing at full res.

Or am I missing something?

[Bill Anderson]

"Or Am I missing something?"
Not yet. (:

[Jannard]

YES! smaller sensor (~1.3x crop) will make it more AFFORDABLE, more sensor per wafer! Making VistaVision's (36mm X 16mm) size sensor definitely increase the cost.
Yes I did, twice! Nice images!! This has nothing to do with 16bits or 12bits. If your photodiode in each pixels is small, then it will have smaller charge collection capacity than a larger one. It acts as a capacitor while exposing the photodiode to light, it's this charge collection capacity that determine the dynamic range of a sensor/photodiode. This is the most fundamental, if the PD can't collect enough photogenerated charges, then you can't have any more dynamic range. If I'm stuck with smaller sensor, then I can't use 16bits. There's not enough number of electrons, or volts, I can represent for one binary number. If I do have a large pixel and I can collect large amount of of photogenerated chrages, then I'm force to use larger bits number to represent the voltage across the PD.

Comparing the images from the Canon 5D and 1Ds MkII is a good example of how much of a difference one micron difference in pixel's size in the accuracy of colors and dynamic range. This is a fair comparison, because they uses the same DigicII image processing!

The new 1D MarkIII produces great images, the fundamentals are all there to begin with.....not to mention the DigicIII processing!

BUT WHAT AM I COMPLAINING FOR!!, after all, it's <$20,000 It's definitely a better replacement of the HVX, etc....

Nanohv... what do you make of Rodney Charters quote that it beats film? And PJ's quote that it compares to 65mm film? We don't see RED as just an HVX replacement.

Jim

[Graeme Nattress]

It's not just pixel size, it's pixel DESIGN that makes the difference :-) Size is just one factor. After that, it's what you do to decode that pixel, analogue circuitry and more.

Graeme

[Graeme Nattress]

Yes, Michael is a friend of mine. He's got a superb site for anyone interested in photography. I think the key thing, that forgets to be spoken is, "All other things being equal", a larger pixel is better than a smaller one.

As pointed out, a larger pixel will, for full saturation have less noise. Although it's not photon noise that is the real problem, but read noise, which really must be minimized, and that's down to design more than size, I think.

The other design aspect related to how much of a pixel's size is active area - ie fill factor. You can have a larger pixel perform worse than a smaller one if the small one has a very high fill factor and the large one a low fill factor.

So, "all other things being equal..."

Graeme

[Rocket]

Graeme, can you expand a bit on fill factor?

[Mark L. Pederson]

I bet the competition is reading this thread and taking notes ...

[Rocket]

Yeah, I was just thinking that myself. Maybe it's not such a good idea to let all the secrets out.

[Graeme Nattress]

All above info is basic know-how for any sensor designer. I'm not a sensor designer, but I do know some of the basics of how things work. As to regards to Mysterium, we are not talking about it's design.

Graeme

[N. Keller]

You can also record 12-bit linear from the Red. It's not a huge difference to 10-bit log - because the log curve puts the bits in the right places.Bruce Allen
www.boacinema.com

Hang on...and excuse my ignorance. Can the RED record directly 10 bit log? I thought it was just R3D, 12-bit linear.

[Andrew M.]

You can flip to 10 bit log in REDCINE, maybe it is what Bruce meant?

Than again REDRAW is compressed 12 bit lin and the way compression works it may as well compress in log fassion since less important information compresses the most.