EX1R - HDMI output 8bit or 10bit?

[Cliff Totten]

Ok, this is a fantastic thread. Thank you to all for the info. I can honestly say that I understand about 81.5% of all that has been said here so far. (Alister can sometimes go over my head a bit though...I've got to read slowly and really pay close attention to keep up with him!)

So, here is my bottom line question;

If you were shooting a rock video in front of a green screen with tons of available light and your goal was to create a very sharp, clean and VERY bright looking video with deep, surreal and over saturated colors...would you shoot negative gain?

From what I'm understanding now, it won't seem to do me any noticeable good. (in the cleanliness of my video...and I'd be trading off some overall dynamic range for it)

I might buy an Atomos Ninja or just capture to a Blackmagic Design Intensity card. (4:2:2, 8 bit it seems)

CT

[Steve Kalle]

The best workflow I have found for grading is to take your recordings, process them with Neat Video (plug-in) and export with Cineform. Then, bring into your program to grade. (I use Premiere Pro CS5 with Neat Video and export to Cineform 422 Film Scan 1 and AE to grade). FYI, if you render to 444 with Cineform, you lose the super whites. I also use a nanoFlash for my high quality recordings (280Mb I-frame).

[Dave Sperling]

Originally Posted by Cliff Totten

If you were shooting a rock video in front of a green screen with tons of available light and your goal was to create a very sharp, clean and VERY bright looking video with deep, surreal and over saturated colors...would you shoot negative gain?

Haven't been shooting rock videos, but I do a lot of green screen and would HIGHLY recommend using an external recorder to give you 4:2:2 color recording.
I personally haven't noticed a big difference in the mattes coming off stuff shot at -3db compared to 0db, but in the last couple of months I've switched cameras and am now using an F3 into my NanoFlash for green screen work. Because I like to use a narrow shutter angle to minimize edge blur, I have been tending to shoot at 0db, and the results have been great.
The other issue in your post is the 'over-saturated colors' -- you may want to be careful because any kind of chroma bleed can be your enemy. If you can do your over-saturating AFTER pulling your key that might be a better approach.
But the bottom line is, I wouldn't consider shooting a green screen at anything less than 4:2:2.

[Cliff Totten]

If I don't pickup a Ninja than I will certainly be capturing to a Black Magic Intensity card using the MJPEG 8bit 4:2:2 codec. (eehh,...not sure about that codec yet)

Another issue with the HDMI out - My favorite frame rate is 29.97p. I love that look over 24p. (I feel that a progressive 29.97p will move nicely into a 60i for Blu Ray if needed...it still maintains the "progressive" look because the fields are sliced from the same frame)

I have a funny feeling that the the EX1r will process 29.97p right up until it gets to the HDMI port. I'll bet that it will break out those frames out into top/bottom fields to maintain TV compatibility. (like the Sony NEX VG10 does)

I plan to be careful with my levels and I'm going to shoot everything within range. My plan is to screw up and exaggerate the colors only afterwards, in post.

Thanks everybody. This has been a great topic.

CT

[Alister Chapman]

Perversely, if you want to shoot something that contains lots of highly saturated colors it can be beneficial to reduce the overall saturation level of the color matrix to prevent colors from clipping in camera and then bring them up again in post. If it gets clipped in camera you will get blocky colors in the finished video. There was a more in depth discussion about this with regard to shooting Hi-Vis jackets in another thread.

[David Heath]

Originally Posted by Alister Chapman

Negative gain reduces dynamic range as it will artificially clip your low key parts of the image. The brightness range is restricted by the cameras DSP, it only has so many bits to sample the sensors output, from dark to light. If you reduce the gain the highlight handling doesn't change (DSP bit depth limited) but your entire image get shifted down and as a result you will clip off some of your shadow and dark information, so your overall dynamic range is also reduced. 3db is half a stop (6db = 1 stop) so -3db gain reduces the dynamic range by half a stop.

That's not my understanding - I agree with the last bit "so -3db gain reduces the dynamic range by half a stop", but surely it's at the expense of highlight handling, not lowlight?

The rationale is that the output of any photosite is physically limited to a maximum value, put more light on it, get a bigger signal out - subject to a maximum.

So assume the output value could increase from 0 to a nominal max value of 1000 (say) for increasing incident light, after that, no matter how much brighter it gets, you don't get a value of more than 1000. Now assume that an output peak white is nominally associated with a value (and light level) corresponding to 200 (say). All that headroom (from 200-1000) is available to allow for processing to give highlight detail instead of a hard clip.

The effect of negative gain is to mean that to correspond to peak white output, the output of the chip must be higher - say 400 for the previous example. So far less headroom (2.5x v 5x) to play with between nominal white and max chip output. Result - less highlight capability, but quieter in the blacks.

[Alister Chapman]

We agree, but are saying it slightly different pedantic ways. For those not interested in the science skip to the end paragraph as that's the only bit that really counts!

With most modern digital cameras it will be the low end thats clipped and the reduction in latitude is a reduction in underexposure range. The sensor itself is a constant gain device and with a CMOS sensor the A/D is done on the sensor chip. Once a certain maximum light threshold is reached the sensor will saturate and cannot output any more. So a fixed maximum dynamic range and bit range will leave the sensor and go into the DSP, the over exposure and under exposure limits of the system are determined by the sensor and it's bit depth or by the number of bits the DSP uses. The DSP will shift the gain up and down whilst applying the gamma curves etc. So if you reduce the gain, it is the bottom end of the available range that will be clipped, as it becomes sub-black. The high end just changes in output value with the signal level reducing for the same amount of light input. More light will not increase the output as the sensor will already be saturated.

Lets consider a scene where we adjust the iris so that the sensor's output is just starting to clip, so adding more light to the scene will not give any greater video output signal.

Imagine your sensor is outputting bit range of 0 to 1000 when the scene goes from black to maximum illumination at 0db, which in turn gives an output range of 0 to 100%. If you use negative gain in the DSP then bit 1000 gets reduced lets say to bit 900 and Black becomes -100 so now what used to exist between bits 0 and 100 can no longer be seen as these are below black (reduction in under exposure range) but nothing changes in the highlights other than a reduction in output level. The dynamic range is shifted down as input bits 100 to 1000 are now represented by output bits 0 to 900 and the output voltage range becomes 0 to 90%. The same illumination now gives only 90% output. This is clearly seen on an EX as the reduction in output level at -3db.

So with our fixed light range scene, the video cameras output will reduce by -3db and the picture will look darker. We will still see all the brighter parts of the scene as at 0db, just a little darker looking on the monitor. What we will not see is what used to be in the shadow areas.

If we do now increase the lighting, as the sensor is already max'd out, we won't see any further increase in the cameras peak output, but then at 0db this would have been in exactly the same situation. So the camera will still handle the same amount of bright light as before but we have lost some of the ability to see into the shadows.

So if you measure form a fixed reference point in the scene then you are loosing the ability to see into shadows relative to your highlights. However, measuring from Black to light it would appear that it is the highlights that have become more limited as the peak output level has been reduced, but actually the amount of bright light the camera can handle hasn't actually changed, it's just become less able to reproduce shadows.

Of course in practice it all means the same thing as you would compensate for the darker image by opening up the iris, which would bring the shadows back up again, but because the sensor is already saturated the peak output won't change.

[David Heath]

Originally Posted by Alister Chapman

Lets consider a scene where we adjust the iris so that the sensor's output is just starting to clip, so adding more light to the scene will not give any greater video output signal.

Imagine your sensor is outputting bit range of 0 to 1000 when the scene goes from black to maximum illumination at 0db, which in turn gives an output range of 0 to 100%. If you use negative gain in the DSP then bit 1000 gets reduced lets say to bit 900 and Black becomes -100 so now what used to exist between bits 0 and 100 can no longer be seen as these are below black (reduction in under exposure range) but nothing changes in the highlights other than a reduction in output level. The dynamic range is shifted down as input bits 100 to 1000 are now represented by output bits 0 to 900 and the output voltage range becomes 0 to 90%. The same illumination now gives only 90% output. This is clearly seen on an EX as the reduction in output level at -3db.

No - gain is a multiplication/division process - not addition/subtraction. What you are describing is an overall level shift - not a gain change. It's also not true to think that max sensor output corresponds to nominal white, in practice it will be very much more.

In the absence of knee etc, if the sensor is starting to clip, it's outputting many times more level than would translate to peak white. Normal working is well below sensor limiting. Let's assume (in the absence of knee, and ignoring gamma) using your own figures that at 0dB peak output white corresponds to 200 as sensor value. If we now have -6dB gain, we need twice the sensor value to give the same output - 400 - which can be achieved by doubling the incident light level, exactly as expected. With all inbetween values also multiplied by 2 - not having a fixed value added.

Conversely, imagine if the light drops to only give a sensor output of 100 corresponding to white, 50 for mid grey (instead of 200 and 100). If we want that to correspond to the previous values, we need to MULTIPLY by 2 - not add 100. Which is the same as adding 6dB of gain.

In your example, it's the sensor values between 200 and 1000 (at 0dB) which normally get processed to give the knee. At -6dB there will be less headroom (400-1000) so the knee can't be as effective.

[Alister Chapman]

I still have to disagree with you David.

I'm well aware that gain is a multiplier. But the application of negative gain will shift a signal range down so my example is completely valid.

With an analog camera I might agree with you, but this is a digital camera. There is an old rule of thumb that says for every stop you want to record you need 1 bit of data, so for 8 stops you need 8 bits of data, 12 bits for 12 stops etc, while technically yes you can record 12 stops with just 1 bit of data (black =0, +12 stops =1) you really don't want to, and the 1 bit 1 stop rule normally holds true before you start applying gamma curves and other compression.

We know the EX1 and EX3 uses a 12 bit DSP, so it's not unreasonable to believe that the sensor is outputting a little under 12 stops of latitude allowing for correction bits etc. within 12 bits of data. This would be a pretty standard system and an efficient use of the available data pipe. Remember a sensor is a linear device, so to add another stop of dynamic range prior to the DSP requires a doubling of the amount of data and this is a budget camera so the designers are not going to use a larger bit depth (with all the associated extra costs) than is necessary to achieve the desired output. Nor do they want to waste valuable data bits with latitude that will never been used. So I hypothesise that in order to maximise the quality of the front end of the camera the designers would probably simply allow anything over 12 stops to clip as digitising anything over 12 stops is an absolute waste of valuable data as it will never see the light of day. So in effect the system is as I said already working within a limited Dynamic Range, probably around 11.5 stops absolute max at the sensor output.

We cannot ignore gamma and the knee, because as has been measured many times the EX1 has around 11 stops of Lattitude with the Cinegammas or a well adjusted knee.

So allowing for some noise reduction and other processing overheads, when you are using a cinegamma or the knee, which is on by default, the camera is likely to be outputting pretty much every bit of available sensor data. Even after subtracting 3db of gain, increasing light levels won't help as the A to D's and the digital pipe into the DSP is saturated due to the 12 bit nature of the pre gamma system.

This is born out in practice by the way the actual output signal from the EX cameras reduces by 3db when you use -3db gain and you cannot, even with more light get back up to full output. The only way to get back to full output is to adjust the gamma curve slope in the picture profiles which increases the gamma gain and allows you to increase the highlight gain. If you do this you can get your highlights back to where they were before you added -3db but still end up with clipped shadow areas. This tells me that the drop in output associated with the use of -3db is not simply a bug in the way the white clip is applied but simply because there is no more latitude to be had.

You can see the very behaviour I have explained with a waveform monitor. If you take an overexposed scene and look at it carefully on a WFM, then reduce the gain by 3db there is no change to the way the overexposure rolls off, it is exactly the same, the highlights are captured in exactly the same manner, just output at a reduced level, but you will see previously visible shadow detail simply disappear as there is now insufficient gain to show those subtle shadow textures, which is also of course why the noise is less visible. Lets not forget that noise is not that dissimilar to dark parts of the scene, normally small random changes in brightness and colour. If reducing the gain is helping us make the noise go away, it is also making subtle dark parts of the image go away, a reduction in underexposure latitude!

Further confirmation of this behaviour can be seen if you use a standard gamma and turn the knee off. Taking the gain down to -3db now does not result in an automatic decrease in peak white output by 3db because now our standard gamma non-knee image only has about a 7.5 stop range (funny that: 8 bits = roughly 8 stops, it's not a coincidence), so now we do have an excess 3 stop of latitude that are not getting used, again this confirms that when using the knee (which is on by default or a cinegamma) we are using the full latitude range available to us, there is no headroom that can be used by compensating with more light. So the sensor/data pipe is limited at approx 11.5 stops.

David, If we had a large excess of headroom as you suggest, then a well adjusted knee slope would give you a greater dynamic range, beyond 11 stops, but this is not the case, with the vast majority of cameras that use a 12 bit data pipe from the sensor to the DSP we hit a brick wall at around 11 to 11.5 stops, it's the point at which the data pipe becomes saturated. It's why as we have seen the number of bit's used in DSP's increase we have seen corresponding increases in dynamic range, it's more than just a coincidence.

So, if you are doing a shoot where dynamic range is an issue, the likelihood is that you would be using a cinegamma or knee to help deal with the wide dynamic range. Reducing the gain by 3db will reduce noise by a small amount but it will also make the darkest parts of your scene disappear. Increasing the lighting to compensate (or opening the iris) for the loss of those shadow areas will bring them back, what you are doing is decreasing your underexposure. However as your highlight range may well already be clipping you will not see any difference to your highlights other than increased clipping. It is very clearly the underexposure range that is reduced and this can be seen on a WFM.

This has been an interesting discussion as it has made me look much more closely at the way the output level of the EX and PMW cameras decreases when you use -3db gain. It's not just a quirk of the way the gain is applied but simply the case that there is no more headroom to be had when using the knee or cinegamma. Auto knee does disguise the effect a little, but a fixed knee or cinegamma clearly illustrates that the camera has an absolute limit of around 11.5 stops coming off the sensor.

[David Heath]

Hmmm - OK, I think if there's any disagreement, it's because I've been talking in very general terms (very idealised and simplified theory) - you've been referring very specifically to how an EX behaves.

In my hypothetical camera, then the blacks should stay at the same level, and all the changes (esp knee slope) should happen at the white end. The settings (knee, gamma near black etc) should all adapt to compensate for the changing gain. When you say "the actual output signal from the EX cameras reduces by 3db when you use -3db gain and you cannot, even with more light get back up to full output. The only way to get back to full output is to adjust the gamma curve slope in the picture profiles ...... then it's obviously not adapting. (Which frankly I think it should be - you shouldn't have to adjust gamma curves manually when you change gain.....)

I also suspect that in the practical case of the EX, it really warrants readjustment of the black gamma with the addition of negative gain. I would still maintain that a camera shouldn't give clipped blacks with negative gain. If it is doing so, it's a sign that it's not automatically readjusting process values near black when the gain is changed - which I feel it should be doing.

Where I won't disagree is in all you say about overall dynamic range - and negative gain shouldn't change that in total. I would expect it to redistribute it (to the gain of the lowlights/expense of highlights), but it sounds as if to reach optimum, the camera needs manual realignment between 0 and -3dB.

Originally Posted by Alister Chapman

it has made me look much more closely at the way the output level of the EX and PMW cameras decreases when you use -3db gain. It's not just a quirk of the way the gain is applied but simply the case that there is no more headroom to be had when using the knee

You see, this is exactly what I meant when I said "it's the sensor values between 200 and 1000 (at 0dB) which normally get processed to give the knee. At -6dB there will be less headroom (400-1000) so the knee can't be as effective. I'd expect a camera to vary the knee processing when negative gain is used - not use the same multiplying factors as for 0dB. Yes, there is less headroom, highlight handling won't be as effective, but the white output level will be what is expected.

Love to talk it all over with you in person, Alister - far too open to misunderstanding to discuss complex issues in internet posts......!

[Alister Chapman]

I think your are missing the point that -3db is not just less gain but it is 3db of attenuation.

The majority of modern cameras exhibit exactly the same behaviour as the EX's, certainly most Pana HVX and HPX cameras do:

Dynamic range is the ability of the camera to handle the tonal range of a scene from shadow to bright light.

The amount of bright light the camera can deal with is limited by the point at which the sensor/data pipe saturates.

The signal output relative to light input is governed by the gain of the camera.

Reducing the gain reduces the relative output signal, but the peak amount of light the sensor and thus camera can cope with does not change. This is a constant, governed by the sensor, so highlight handling does not change.

Attenuating (-3d is an attenuation, not just less gain but negative gain) small signals may make them so small that they are no longer visible, Signals that are only 3db above zero become zero. That's why the noise level decreases.

Shadow areas of the scene are small signals. Attenuating these by 3db may make these no longer visible. Parts of the scene that previously were 3db above zero are no longer visible.

But the upper limit of what the sensor can handle does not change, only the relative output signal.

So: Attenuating the incoming signal from the sensor by 3db will prevent the bottom 3db of the sensors signal range from being seen while the upper range does not change, only the relative output level.

Using 3db of attenuation reduces the shadow performance of the system (small signals near black are lost, but bright light (and thus overexposure performance) does not change. So dynamic range reduces by 3db due to a loss of shadow information.

See attached graph, or see Art Adams explaination here: ProVideo Coalition.com: Stunning Good Looks by Art Adams

[Cliff Totten]

What a fantastic discussion. I hope more people notice this tread and take a look at it.

Alister, I dont think it's possible to explain this the technicals of situation and better than you just did. For a while, I often wondered why cameras didn't have deeper negative gains than they do. (if -3db would reduce noise than why not -6 or -9db? enabled. etc...) It's now clear to see why!

I will be shooting at 0db for sure. Bottom line for me is that I love dynamic range more than I hate noise and the trade off value doesn't seem worth it.

Thanks again to Alister and David.

CT

[David Heath]

Originally Posted by Alister Chapman

Using 3db of attenuation reduces the shadow performance of the system (small signals near black are lost, but bright light (and thus overexposure performance) does not change. So dynamic range reduces by 3db due to a loss of shadow information.

See attached graph, or see Art Adams explaination here: ProVideo Coalition.com: Stunning Good Looks by Art Adams

I'm afraid that Art Adams explanation is fundamentally flawed. With reference to his diagrams, he says:

The sensor sees what it sees, and that’s it. The image we see through the viewfinder can be thought of as a “window” through which we view data collected by the sensor. The gain switch moves that window up or down the sensor......

The first sentence is completely true. The raw sensor output will depend on incident light, end of story. But the "window" (which represents the subsequent processing) analogy he gives is wrong. Simply moving it side to side in his analogy represents level shifting - not gain changing. Gain changing would correspond to making the window smaller or larger.

Think of it this way. Assume the raw sensor output can go from 0 (black), to 1000 (bright as it can be) - and these numbers are ONLY dependent on incident light. Now assume that for "0dB processing" they are translated into values 0 to 100 - so 0 output = black, 100 output = white. 10 input values map to 1 output value, OK? Now, the incident light level is halved - so the sensor values only go from 0 to 500. We would expect 6dB of gain to give us back a max output of 100 - the gain compensating for the halving of incident light - so in this case the 500 must be made to map to 1000. As far as processing is concerned, the impact of the 6dB gain increase is to mean that now 5 input values map to 1 output value, not 10:1.

In Art's "window" analogy, the window at 6dB is now only half the width it was at 0dB. The window has changed in size, NOT just moved, as he says.

In real life, gamma, knee etc make it far more difficult. but they only really represent a dynamically changing gain with light input level. Normally highest at low light levels, "nominal" in the middle of the curve, and low in the highlights. Practically, the variation is continuous, but let's assume it happens in three discreet parts for ease of explanation.

(Some readers may wish to stop here....... :-) )

Still using the 0-1000 model for raw sensor output, and 0-100 for output, let's assume that the lowlights are 0-10 in terms of output, mid range is 10-90, and highlights are 90-100. We'll assume that "normal" 0dB gain represents a 5:1 mapping in the mid tones, and 2:1 mapping (higher gain) in the lowlights. Hence, input values 0-20 map to 0-10, 21-420 map to 10-90, and 421-1000 map to 90-100 (about 60:1 - far lower gain).

Now let's assume we reduce the gain by -6dB (numbers are easier than -3dB!). The low end mapping becomes 4:1, mid-end 10:1. But IF we now map 0-40 to 0-10, and 41-840 to 11-90 AND open the iris a stop, the effect on the low and mid-tones (all the way up to 90IRE) is exactly the same. Dropping the gain does not necessarily mean a lack of shadow detail - in this case it will be exactly the same, and with lower noise due to the lower gain near black..

Of course, in that example there would be a terrible price to pay in the highlights. Opening the iris a stop would mean that many output values would be saturated compared to the first case (give a value of 1000) and highlight detail lost - but that's exactly the point. Negative gain can (OK, should) improve performance in the blacks - at the expense of highlights. You're robbing Peter to pay Paul.

That is obviously an extreme example, and in real life a lot of things will be different. But it begs the question of why the EX and Panasonic cameras mentioned by Alister and Art behave as they do. I suspect it's because the differential gain changing doesn't change with overall gain as it ideally should. In the above example, it corresponds to applying the high lowlight gain to sensor values 0-20 (as for 0dB) - rather than modifying it to 0-40. As such, it's right to see it as a bug, and not something that is inherent.

[Alister Chapman]

Originally Posted by David Heath

I'm afraid that Art Adams explanation is fundamentally flawed. With reference to his diagrams, he says:

The first sentence is completely true. The raw sensor output will depend on incident light, end of story. But the "window" (which represents the subsequent processing) analogy he gives is wrong. Simply moving it side to side in his analogy represents level shifting - not gain changing. Gain changing would correspond to making the window smaller or larger.

But you are ignoring the fact that 0db does not mean zero sensor output. 0db is likely to be some way up the sensors output range at the point where the engineers have decided that the signal to noise ratio represents the optimum compromise with regard to sensitivity and DR. Then you have to remember that -3d is not just a reduction in gain but an ATTENUATION of the input signal.

Originally Posted by David Heath

Now let's assume we reduce the gain by -6dB (numbers are easier than -3dB!). The low end mapping becomes 4:1, mid-end 10:1. But IF we now map 0-40 to 0-10, and 41-840 to 11-90 AND open the iris a stop, the effect on the low and mid-tones (all the way up to 90IRE) is exactly the same. Dropping the gain does not necessarily mean a lack of shadow detail - in this case it will be exactly the same, and with lower noise due to the lower gain near black..

Of course, in that example there would be a terrible price to pay in the highlights. Opening the iris a stop would mean that many output values would be saturated compared to the first case (give a value of 1000) and highlight detail lost - but that's exactly the point. Negative gain can (OK, should) improve performance in the blacks - at the expense of highlights. You're robbing Peter to pay Paul.

But that's exactly mine and Arts point. To compensate for the lower gain you have had to open the iris to regain your lost/reduced blacks and mids. The highlight handling of the sensor however has not changed, so now your highlights have become over exposed due to the need to open the iris a stop to regain the lost low end. Your having to deliberately over expose to compensate for lost low end, but because your over exposing your highlights are now blown out. If you didn't open up the iris the highlight handling would remain the same, so highlight performance has NOT changed, it's governed by the DSP's data pipe. And what happens to the input signal from the sensor that is now being ATTENUATED by 6db? Zero sensor output does not equal zero signal.

You muddling signal output levels with the actual dynamic range of the camera. Differing amounts of positive gain will of course vary the output signal level, possibly taking it above or below peak white, but this is an output level change, not a dynamic range change. ATTENUATING the input however will mean that some small signals will be lost and this results in a change in dynamic range.

Originally Posted by David Heath

But it begs the question of why the EX and Panasonic cameras mentioned by Alister and Art behave as they do. I suspect it's because the differential gain changing doesn't change with overall gain as it ideally should. In the above example, it corresponds to applying the high lowlight gain to sensor values 0-20 (as for 0dB) - rather than modifying it to 0-40. As such, it's right to see it as a bug, and not something that is inherent.

I think my graph explains the behaviour of almost every current Sony and Pana camera. You have a limited optimum signal range dictated by the DSP, especially at the high end where every additional stop of over exposure requires a doubling of the amount of data used (remember a sensor and it's data output is linear). You cannot ever exceed that maximum data point, so the highlight handling remains fixed. The zero point will be somewhere above zero sensor output to allow for noise reduction and noise floor/sensitivity optimisation (and outputting underexposure requires a fraction of the data as over exposure). Attenuating the signal by 3db will reduce the maximum output level by 3db, but as this is well above zero this ends up as nothing more than an output signal level change (as seen in the 3db peak output drop). But attenuating a small signal that may only be 3db above the noise floor by 3db will mean that small signal is now zero and is lost.

If you attenuate the input signal going into an amplifier by 3db you won't see or hear signals until they are greater than +3db, so therefore they are lost, while signals greater than +3db will be heard or visible but now at a reduced level. So, low end is lost, high end the same but reduced level. -3d is an ATTENUATOR. not just less amplifier gain.