High Dynamic Range?
 

In reading the benefits of the CMOS technology, one rapidly encounters the term "high dynamic range" as one of the benefits of CMOS over CCDs.

It seems to me that a lot of the camera reviews and tests out there put a lot of energy into estimating the low-light performance of cameras (a known challenge for CMOS), but I was wondering if there were any comments, images, or tests on the HC1 that validate the claim by Sony that the CMOS chip offers higher dynamic range.

-Steve

Lmiting factor is recorder's 8 bit dynamic range. FX/Z1 CCD dynamic range many many times higher. CMOS even higher dynamic range but for practical purposes is meaningless because you just use ND filters with the CCD camera.

Radek

No - 8-bit only limits the number of samples from light to dark to 256 per channel... it does not define the spacing between those intervals, and hence has nothing to do with the dynamic range (from light to dark) of an image - only the number of possible steps. What I'm talking about is often referred to as lattitude - and is related to contrast, saturation and gamma.

-Steve

To Steve: I think in Kaku's post, Wayne Morellini said something about how the HC1 exposes pixels individually to provide proper exposure of the image. You should check out his post for the details.

To Radek: As far as I know, ND filters *do not* increase a camera's dynamic range. ND filters reduces the amount of light that enters the lens, so it darkens the whole image; preserving the range of brightness to darkness.

Like Steven mentions, dynamic range hase nothing to do with a/d conversion. The other thing to remember is that dynamic range doesn't tell anything about sensor sensitivity. It only defines a gayscale range from the (noise limited) lowest light levels in a scene up to the highest (saturation limited)levels.

Radek, I have nothing against you but based on the number of times you have posted incorrect information here, I think it would be better if you did a lot more reading/studying about video and a lot less posting.

I have to stand up for Radek. This sentence is not quite true:
8-bit per colour channel is the most limiting factor of dynamic range. It means only 256 shades are possible, you can set diffrent spacing between steps but this won't allow to display high dynamic range scene, because details will be greatly losts. CCD probably has today internal 12-14 bits range, we now have to compress it to 8-bit. There are some basic algorithms:
- linear compression (mostly seen in todays camcorders) causes that shadows are mostly black
- logarithm mapping - causes loss of contrast in highlights but shadows remain details
- local operators - the best one, retains details in highlights and shadows, but can cause halos around edges. It's very complex, HDV image (1440x1080) needs 1 seconds for one frame on todays PC (2GHz). Probably CMOS can be used to implement it in hardware, so I supose that's why this topic arise.

English is not native language for Radek and the same for me, so please a bit more forbearance.

BTW Kaku Ito already posted Bamboo clip that deals with dynamic range and there is no diffrence comapring to FX1. There is also Ban clip that shows what HC1 can achive: you see what's inside car but also almost no detail loss for outside bright environment.

I'm close getting MA in film, from one of best film schools in Europe, have worked in film, TV, including 35 mm, HD productions, have FX1E.

There are three kinds people, ones with ideas, which can sometimes be wrong, ones with no ideas, they can't be wrong. Then are people that will promote DVX or HD1 over Sony Z1. One can pretty much figure why.

Let's look at lattitude this way. Footcandle level on sunny day can be up to about 10,000. Camera has 32x ND filter. With that camera will function fine in brightest sunlight, without closing lens down. 32x ND filter mean reduction lighting level from 10,000 to 300 FC.

The camera will also work in 3 lux level, or 0.3 FC. 0.3 lux would be lighting on brightest area. Lighting in darkest area would be small fraction of.

Contrast ratio of scene will be much much lower than range of CCD. Will be limited by tape, not sensor.

Scene can be lighted with contrast ratio in tens, while sensor will work well in many times higher contrast ratio.

Radek

That statement presumes some kind of universal superiority of one particular model or manufacturer over all others. To push this sort of personal agenda consitutes a platform war. I will not allow that sort of highly undesirable behavior on my message boards. It simply will not happen here.

Technical inaccuracy is one thing, and it's forgiveable. Hidden agendas, however, are not. Radek, you'll need to contact me by email. You and I will have a private conference before I'll allow you to post here again. Thanks in advance,

This kind of thing can rapidly approach semantics... but I think it's fair to argue, that 99% of the digital images we see are limited to 24-bit (8-bit per channel) colour, and that there's a sufficiently wide range of possible contrast ratios, gamma curves that we can safely state 8-bit isn't the limitation on dynamic range... it's how those 8-bits are used. Going to 10-bit or 14-bit will most certainly improve the "quality" of the image by adding detail... but if you define "white=x and black=y" to fundamentally the same analog signal, if you continually increase the bit depth, at some point you'll just be measuring noise more accurately.

The lack of an ND filter on the HC1, and the "lack of necessity" of one, seems to clearly point to a very high range on the sensor, and the dynamic range of the actual footage will depend largely on how Sony implemented the controls. Considering the limitations of manual control on the HC1, I imagine the full utility of the sensor can barely be realized.

Perhaps in the future someone will implement a consumer CMOS camcorder with fully adjustable dynamic range and gamma. Wouldn't that be something?

I'll have to find that clip/shot.

-Steve

Michael, dynamic range is a property belonging to the analog world. Scene contrast ( real world dynamic range) on a sunny day is much higher than cam sensors can get. CMOS is the better device in thet respect. The number of bits used for digitizing analog values is rtelated to the target application. For video it is generally accepted that 8-bit depth resolution per color doesn't show contouring if applied in a overall linear chain. Signals from cam sensors need to be preprocessed (gamma correction, WB) and therefore more initial depth is needed, like 12bit and more. The processed values are then remapped on an 8 bit structure (interpolation).

I can't agree that 8-bit (256 shades) is enough to reproduce dynamic range (or high contrast, whatever you call it). I know the contrast ratio on a sunny day is much higher than cam sensor. But it's NOT possible to simply put this in 8-bits, be happy and say that 8-bit is enough and doesn't have anything to do with dynamic range. If we have contrast ratio of 10 000 : 1 (some of newest displays allow it) but still 8-bits then we stuck to nothing - the true scene with 10 000 : 1 will look on such display increadibly unnatural.
Just look here:
http://www.cs.huji.ac.il/~danix/hdr/hdrc.pdf
http://www.mpi-inf.mpg.de/resources/...map/logmap.pdf
If it would be so simple as you describe then above algorithms won't be necessary. And these algorithms can't be implemented real-time. Feel free to ask if you don't understand something in those papers.

8bit is only good enough to ensure lack of banding after gamma correction. If the camera were to record linear video, it would need to be at 12bits to eliminate banding.

Graeme

Michael, thanks, and be shure I do understand yr stuff. Already duing the time I was involved at Stanford University(CA) in medical image processing about 30 years ago and later, till 2 years ago as VP R&D at Barco (900+engineers in research) I was supposed to know where I was talking about. So no need for extra leads, maybe I could send you some of my reports and patents on image processing, including one on histogram optimisation (about yr local operators) for medical ultrasound imaging.

Andre, I was mostly replying to Steve post. You've written about contouring etc. and that's ok. So maybe we are floating at the barrier of my communication (my English is not perfect ;) ). I agree that 8-bits is enough to represent accurately image, BUT we need a bit diffrent capture devices that doesn't cause contrast compression but work more like our eye. Today available professional digital camcorders with ultra high dynamic range meets the problem how to compress the contrast to only 8-bit. Maybe with CMOS we are closer to this (human eye) approach and soon maybe we would not need to do any compression of dynamic range and the image will be displayed perfectly on our limited monitors/TVs.

Agreed Michal.

But are you sure the camera doesn't have internal undocumented ND filters like the PDX-10, TRV-950, HC-1000 and many other cameras? This design lets the camera stay in the sweet spot of the lens without ever using small apertures. Instead, one of several internal ND filters drops into the optical path.

The existence of these internal filters have been proven on the above mentioned cameras, but Sony has never published anything about them and the user has no manual control of their behavior.

There's a difference between not documenting something and saying something definitively isn't there.

I never said that it was. 8-bit is both a display limitation and a recording bandwidth limitation in the already highly compressed land of digital video. And before you quote 10000:1 contrast ratios on televisions, unless the television can take a standard input signall that supports 10000:1 - however that's tested - and put that on screen, such figures do little except tell you how black the blacks are compared to the whites on display - not in the signal supplied to the display.

The reason I'm even curious is becuase the game with camera dynamic range is to record as much useful information as possible - and is the same game we play with everything from gain to iris to shutter speed. Having a large dynamic range would be beneficial while shooting high contrast scenes, to enable more control over shadow detail and highlights for future colour correction. If it can't be adjusted manually though -it's all for nothing. I'd love to be able to carefully set white points, black points, and everything in between in camera.

And André! Cheers to a fellow in image processing. I did my M.Sc. in nuclear magnetic resonance imaging, and am currently playing with non-destructive testing methods.

-Steve

Sorry, I thought you've said that, so it must be my english.
Wow! That's really interesting!

Hm... it's probably derived from this statement:

"there's a sufficiently wide range of possible contrast ratios, gamma curves that we can safely state 8-bit isn't the limitation on dynamic range"

What I meant to indicate is that pretty much all the digital video and pictures you see on the web, TV, etc. are 8-bits per channel. People seem to be happy to differentiate - within all these 8-bit delivery formats - between scenes with high and low contrast, scenes with and without blown out highlights, etc. So the 8-bit isn't so much a limitation on the presentation... it's a limitation on the colour correction flexibility.

You've got all these people claiming some images scream "video" while others don't... yet they're all looking at 8-bit pictures. I argue 8-bit delivery isn't the problem - it's where these 8-bits come from. The more you can adjust gamma, and the values of black and white, the more you'll be able to specify the dynamic range for the image you want. Logarithmic, polynomial, inverted or sinusoidal for all I care... not to make the images "natural" but to make them what the "artist intends".

Getting what you want is in principle very easy from high dynamic range source... doing it from low dynamic range source isn't. In the HC1 you have a high dynamic range CMOS sensor and you've also got a 14-bit signal processor. I would expect there's a lot of flexibility in what the camera can pick out of those 14-bits before downsampling to an 8-bit export - but whether the folk who built the camera let you get at it, or whether they decide everything for you and pre-set it is the real question.

-Steve

Steven, I always found image reconstruction from NMR data so challenging...much more sophisticated than what is being done in videoland. Is your NDT work still MR related? I always was interested to know how the MR related techniques like NQR used for NDT would affect our magnetic media when used in airports.

Totally off topic - but I really marvel at the genius of the NMR encoding schemes. I tip my hat to all the folk who came up with it originally - truly an elegant procedure.

Not really... I'm working on detecting stresses in steel pipes using DC and AC magnetic techniques (known as flux leakage and Barkhausen noise respectively). It's still related to magnetism, and the work is for the nuclear industry... but spin polarizations and gyromagnetic ratios aren't so much a concern. Most ferromagnetic effects are due to electron dipole moment coupling.

I'd never heard of this until today (for shame!). After looking it up, I don't think it would have much effect. The QMR people don't seem to need a magnet and are operating with radio frequencies from MHz to GHz. I think they'd need pretty high power pulses to do much damage to magnetic media. Also, they're only trying to get a signal out of the samples - they're not trying to heat them up (think a microwave oven), so it's doubtful they'd do much damage. According to this article:
http://gazette.gmu.edu/articles/index.php?id=4925
it's all very safe. Notably amusing is the passage about the RF doing "damage" to "airport workers". Media terrorism at work. RF is non-ionizing radiation, and you're bathed in it all the time. Think cell phones, radio, and well, broadcast TV.

Ah! There we go... back on topic ;)

-Steve

Thanks Steve

I keep a reasonably open mind to Radek (and virtually anybody that has a reasonable attitude). I don't know everything, and can learn things off of him. I wouldn't even mind working with him on a job. We all start somewhere and learn more and more. To me it is about the exploring and learning of accurate information, of use somewhere. I've learned much over the last couple of years here (that I wish they taught at media school, would have been good help for picking equipment) and my head is aching just looking at some more of the stuff revealed in this thread ;) We can state things with a qualify like "maybe" and/or ask for verification or questions to better understand things, so those who do known can enlighten us, and reason things out. But it is upto the stater/questioner to determine, sort through, and receive the correct answers. I don't mind a bit of debate, as long as it is about determining the truth.

Radek,

I'll do a summary of what I think I know (qualifier statement), please disregard all my previous statements, I think this is more accurate (any comments, anybody?). Latitude proper, is the range of light the sensor can record, and the bit range determines the fineness this range of latitude canbe divided up into. Even though the bits don't cause actual reduction in latitude, they do cause usable stops below the minimum value to be lost. For instance (using low values) say the sensor pad well mechanism has a capacity of 1024 electron volts (ev) and after 8 bit conversion the minimum value is 4, to stops maybe lost). If it had a capacity of 256 ev, then the minimum value would be 1, but no usable stops are there to be lost. Note the latitude did not change, the one with bigger well capacity still had 4 times more latitude, even in the smallest value. So yes, you can get more stops, but only if they were lost in the first place. Sorry to use the specific technical terms, but there are a number of similar confusing terms surrounding this.

Another thing that clips the lowest usable stops is noise, normally called Signal to Noise ratio. Noise is produced by many circuits in the camera from sensor mechanism to A/D converter to D/A and analogue output connector. So you can only get more than 8-bits effectively usable if signal to noise ratio is high enough to let you (approx 60+bits for 10 bits) it can go so high (especially using gain) to wipe out the lower bits of 8 bit video.

These might be the internal technical measurements of a sensor chip, but they are not the ones used in tests, as lens systems, Quantum Efficiency of light conversion and Fill-factor on the sensor also effect the outcome.

Lens aperture canbe used to compress light range down to fit the well capacity of the camera (increasing latitude of the shot the camera can deal with) but at the expense of increasing the DOF. To stop this you use ND filters to reduce the light range, but I would imagine that maybe the pass through response curve is bias against the bottom end, producing some clipping there.

Quantum Efficiency, is the concession rate of photons to usable electron volts, 10% QE requires 4 times more light than 40% QE, so given the same well capacity 10% works in two stops brighter light (though probably requires two stops more light to overcome noise). Similarly, the fill factor (the amount of the sensor surface that actually collects light) also determines the amount of light (on downside the missed light increases noise).

Kin

I suspect this is the case, sample shots look like it, but I don't know for sure. I have asked if somebody could read that section of the brochure in Japanese to verify, but didn't get an answer. The dual slope and gain changing techniques are really other tricks to get higher latitude. One resets the pixel and re-samples to get more range, and the other applies gain (from what I can tell) to achieve similar affect.

More answers to Dynamic Range found in the exposure settings
 

Being curious about how the exposure control works on my HC1, I recorded a section of tape cycling through all 24 different settings in sequence. Then when playing back I turned on the "Data Code" feature to see what aperture and gain were being chosen for each setting. From that I made up a chart. Interesting was that across a broad range of the 24 exposure "notches", the aperture and gain settings are the same! A total of 8 notches all have f4 and 0dB gain! How can this be since that part of the range definitely modifies exposure in a noticeable way? After pondering over it for a bit, I realized that just like you folks have been pondering here, there must be another third parameter getting tweaked, a choice of what range of bits to grab out of the 14 bits coming out of the CMOS sensor! Here is a write-up of what I think Sony did to implement their exposure control to obtain high dynamic range from the CMOS sensor. It's a little long, and includes a chart, so I decided to put it on its own page:

http://hdvforever.com/hdv/exposure

There is a fair degree of speculation regarding which specific bit ranges are used in which specific exposure settings, and some of the marketing literature on EIP hints at the process being a bit more complex than what I've described here, but based on everybody's musings I expect something like this was used in the HC1 to achieve its excellent dynamic range it offers in well-lit settings.

Hope you enjoy the read.

-Lorin

A 5 bits window at the highest gain setting is really low...besides the patchy noise also banding must be horrrible, or is it masked by the noise? Gamma corrections arlready takes about two bits,maybe including WB. Are you shure there is no analog gain change before ADC in these low level situations?

Interesting Lorin!

The only thing I'd hold contention with is your "bit-range" description.

You don't "pick" 3-14 of the 14 possible bits, you store each number on a 14-bit scale. That is to say, each channel has 16384 possible shades from black to white. You then process the colour, and down sample to 8-bit (256 shades) for final rendering.

In other words, you sample each pixel with 14-bit precision - what you vary between settings is the voltage you assign to 1 and the voltage you assign to 16384. So if you have a possible range of 0 V to 5 V, in one case you sample 0 to 1 V, the next case 1 V to 2 V, the next range 2 V to 3 V, etc.

-Steve

Lorin,

Isn't the camera adjusting the shutter speed for that mid part of the range? If they don't go lower than 1/30th for the low light (unless you specifically tell it to) they have the whole range up to 1/10000th. There are 8 stops between notch 5 and notch 13 which is a range of 256:1 - just about exactly the difference between 1/30th and 1/10000th (333:1)

I cannot see how they would be playing with the 14 bits in the way you suggest - that would remove all the flexibility to do the shadow/highlight DSP that Sony claims!

Gerald

Lorin,

The 8 settings to the left of the exposure bar on the HC1000, TRV950 and PDX10 are where the ND filters come into play. My guess is the HC1 DOES have undocumented ND filters built in. The camera stops at around an f-stop of 4 and then starts using ND filters, which the camera cannot accurately measure. That's why it doesn't go any higher than 4(I can get my HC1000s to 4.8). There's a way to see if the camera is using ND filters by placing it in AUTO EXPOSURE and shining a mini MAGLITE down into the lens while watching to see if you notice any little blades flip into view way back in the back of the camera. These are the ND filters. Give it a try and get back to us!

P.S. Make sure it is on manual shutter!

Geez, after reading all this techincal stuff, I need an aspirin. hehehe

I thought modern CCDs and CMOS chips basically processed their images at higher than 8bit, than dithered down to 8bit (different than compression BTW). There is a whole science and business surrounding color dithering (fooling the eye into seeing more color than is actually there in the end result)
Many patents are involved in doing this, and it sort of explains why the expensive rigs like Davinci and Flame get away with charging more, since they offer the option of working in higher than 8bit color.

What is really weird is the graphics card manufactureres could give us more color, but wont. Chicken and egg stuff I guess.

Joe I don't think that dithering is involved in the 10 (or more) bit to 8 bit conversion. This is being done through a HW lookup table (LUT). Spacial/temporal halftoning (dithering) is more related to output signal processing.

Stephen and Lorin,

I've been doing a little experimenting with my HC1, comparing it with my old GL1. I took video in all sorts of lighting situations from a fairly dark indoors to bright sunlight and drew up a table comparing the data codes between the 2 cameras on auto exposure. Max gain of the amplifier on the darkest subject was 18dB. Both cameras use 1/60sec as min shutter speed. The GL1 has a manual 4X (I checked the ratio) ND filter which was used in the brightest scenes but I took account of that in the calculations.

What I found was that as you go from darkest subjects to brightest subjects there is about a 2.5 stop (6X) automatic change in the HC1 sensitivity between the point at which the amplifier gain gets to 0dB (very shaded outside scene) and a bright outside scene (not the brightest). At the same time as the sensitivity is changing the aperture goes from f1.8 to f4.8 and the shutter speed goes from 1/60 to 1/125. Further brightening of the subject is then compensated by a faster shutter speed - up to 1/250 in my case.

Thus there does not seem to be any doubt that something is giving the change in sensitivity. I have looked in the lens as Stephen suggested but see no ND filter flipping down. In any case I found that the sensitivity change is not sudden but gradual, meaning that there is some sort of 6X (only 2.5 bits) gain change in the DSP. It is probable that with 14 bit resolution in the AD converter they are giving up this 2.5 bits to add to the range of useable light level.

The lens can only stop down from f1.8 to f4.8 (about 7X light range) - it's only a 2-blade control.
The amplifier is 18dB (8X)
"Bit adjustment" gives 6X
The shutter (in my case) went from 1/60 to 1/250 (4X)

This gave a total adjustment range of 7X8X6X4 = 1344 (just over 10 stops)
although more is possible for extremely bright scenes by using a faster shutter.

I hope that this analysis is useful and makes more light than heat or fog!!!

Gerald

Adding a little to my diatribe yesterday.....

I can now see why it is that Sony does not give direct manual control over aperture. The range they have chosen to use is only from f1.8 to f4.8 - not giving a lot of flexibility in depth-of-field etc that normally come from that control. The limited aperture range however is beneficial from the point of view of the optics design - I'm sure it makes it easier for Zeiss to get the best definition over the whole telephoto range. My GL2 goes all the way from f1.6 to f11 which is a much more useful range but the resolution required is much lower.

Does anyone know what the range of aperture is on the FX1/Z1?

Gerald

Higher f-numbers for those small sensor structures would kill HDV resolution due to diffraction effects. Even when the ideal (=diffraction limited) optics would be used.

So the HC1 uses shutter speed to attenuate light past f-stop 4.8? UGH!

The cam itself claims f8...
 

Try this test: Set the shutter speed at 60, film a sequence including bright light, experiment with all 24 settings of the exposure, and then play it back with the "Advanced Data Code" on. The reading in the lower-right corner of the screen claims that it stops it down to f8. Each of those 24 settings became the basis of this article I wrote for mywebsite:

http://hdvforever.com/hdv/exposure

-Lorin

Lorin,

Obviously my tests were a little misleading. I did not try fixing the shutter at 1/60 for my main set of tests - I was using the full auto mode - wanted to see what happened to aperture, gain and shutter. In that mode the aperture stopped at f4.8 and then the shutter speed was varied for brighter conditions. Maybe they are trying not to get too much depth of field, or maybe the definition suffers slightly at smaller apertures due to diffraction.

Anyway, I have tried your "24 steps with the shutter fixed" and I actually got a slightly different result. You found 8 steps at f4 and a minimum aperture of f8. I found 7 steps at f4 and a minimum aperture of f9.6 (then "Closed")!!!

So we can now say that the min aperture varies with the camera manufacturing run - either f8 or f9.6. It remains to be seen which will end up as the final choice when they are in full production.

I think you are probably close on your bit range adjustment. They must use something like this to give the 12X change in gain required for the 7 steps (3.5bits) at f4 (8X or 3 bits in my case). I see that you have presumed that the MSB is only adjusted by 2 bits with much more significant changes in the LSB. This would lead to the picture getting a little dimmer at low light levels (which it does). Good work!

Gerald

Thanks Lorin, i try my HC1 E
 

Lorin, thnaks for ur info, I mange to get a very good exposure control on HC1E
be short, first set shutter to 1/50s. well u may set to any as u prefer. then use exposure notch to get max iris. 1.8 but not gain. simply move the notch to far right and back 6 notch.

so now u get the max iris and no gain to shoot better DOF.
if u prefer f1.8 to shoot , when it's too bright, iu may need to add a ND filter.

i use zebra to get my reading and refer as IRE or zones.
again i find if u set at zebra at 70 IRE , it's about 3 notch plus to 100 IRE or 3 notch down to 50 IRE.

use the histogram to see ur zones, it's very handy too.
i just remeber how it go whne i reading one tone, ( says a grey card or white paper) and set to the IRE i want.

to my own taste i when i use auto i set the AE to -4 and i find the reading is matching with 50 IRE.

so in a quick set up, may just set AE -4 and do the reading then lock it or ajust to what u like.

I, find the dynamic range is quite OK on Cmos.

enjoy.
J.M.