Well I’ve been out and shot a few short clips with the Gemini just to shake it down before I fly to Mumbai in the morning. It works well, even though the firmware is a very early development issue. First thing that got me is just how bright the screen is. Look at the picture here and compare it to the not too shabby F3 LCD!
I have a free day on Tuesday to shoot some local footage in Mumbai, which I will be doing with S-Log and the Gemini. I’ll get some frames on;one as soon as I can.
The question over whether to deliberately underexpose or not with S-Log came up recently. I believe that you need to evaluate the entire scene when shooting S-Log and that the often heard “underexpose by a stop” methodology may have some issues. Here’s my take on the situation:
A couple of caveat’s first: Most of my F3 S-Log work has been in indoor situations as I have been tied to recording to various less than portable 10 bit recording solutions, so very often using a restricted contrast range. I’ve only owned S-Log for my F3’s for a short while now, so many of my earlier tests were on 3rd party cameras, some of these were beta cameras.
I have not fully tied down my workflow. I’m still investigating external recorders, everything from the Ninja, Ki-Pro, Sound Devices and of course Gemini. I’m leaning very heavily towards the Gemini as I do a lot of 3D and the Gemini LCD makes for a fantastic monitor.
Back to exposure, this is obviously going to be a slightly contentious area as there is no real “correct way to do it”. While I might not agree with pinning skin tones or anything else for that matter to one particular brightness range, that does not mean I’m right and anyone else is wrong, it is just a different approach and methodology. At the end of the day, if it works for you and gets the results you want, then that will be the way you should go, these things are not black and white, right or wrong.
A very un-scientific test that a did a while back was an eye opener for me. I was exploring the finite latitude of S-Log compared to the F3’s cinegammas. I did a couple of very quick shots, you will find them here: http://www.xdcam-user.com/2011/06/pmw-f3-s-log-and-cinegamma-quick-look/
When I filmed these two examples I was looking at dynamic range, I exposed in both cases with the bright whites of the back wall behind the girl just going into clipping so I could then see how far into the shadows I could still see useable detail. I was not concerned about getting the skin tone exposure correct. When you look at the raw S-Log it really looks pretty shocking and even I wasn’t sure how much I would recover from the highlights and the girl is a good stop overexposed. However after a very simple grade using only the colour corrector in FCP, I was able to extract a pretty good looking image and it’s amazing how much detail was actually retained in what looked like over exposed high lights. The Girl’s skin tones which I’ve measured at over 85IRE came down very nicely without any issue. A proper grade in a grading suite would I’m sure improve them still further.
What this very crude test told me was that you have incredible flexibility over where you put skin tones, you can comfortably move them up and down in post by a quite significant margin. Also seemingly overexposed S-Log highlights will contain surprisingly large amounts of fully recoverable detail. In the same test I graded the Cinegamma material to try to recover the shadow detail that was lost by due to the reduced latitude. This involved attempting to pull up the shadow areas. While this was somewhat successful, what became very apparent was the way the noise increased quite dramatically, this is something I have been aware of since I started using Cinegammas many years ago, pulling levels up will increase noise.
So… when I expose with Cinegammas (as I have done for many years) I have always been very conscious of the noticeable effect on noise that trying to lift underexposed parts of the image has. Very often in the grade the limiting factor as to how far you can push the image has been down to the noise floor and noise effects. This has mainly been with Sony EX’s which have a 54db noise floor.
Now with the F3 with have a dilemma! S-Log gives us another +1.5ish stops of dynamic range, but at the expense of a +6db increase in noise due to the +1 stop increase in sensitivity associated with S-Log.
Lets say for example that we shoot a shot with a person and we under expose the face by one stop (one stop = 6db).
If we do this with with the Cinegammas and then grade the shot bringing the face up one stop then the noise will increase by 6db from the base noise figure of 63db giving a final noise figure of approx 57db (in the case of signal to noise, a lower number is worse).
If we do this with S-Log and then grade the shot bringing up the face by one stop then the noise will increase by 6db from the base of 57db giving a final noise figure of approx 51db.
So the S-Log image becomes twice as noisy as the cinegamma material and therefore depending on the footage, it is quite possible that you would actually be able to push mid ranges and shadows further with Cinegammas than S-log in an underexposed situation due to noise issues. The S-Log and Cinegamma curves are almost identical up to over 50IRE, so latitude performance under 50IRE is essentially the same. See the charts on this page: http://www.xdcam-user.com/2011/05/s-log-a-further-in-depth-look/
If I get some time at IBC I might see if I can set up some tests to show this in practice.
Now given that I have seen for myself how with S-Log skin tones can be pushed down just as much as up in post, I tend to try to evaluate the entire scene and consider how it will be treated in post before choosing how to expose. In particular I don’t want to expose so that the entire scene will end up being lifted by a significant amount, as noise will become a concern. This isn’t always going to be possible as there are many shots where highlights have to be protected, but I don’t believe that you have to set skins etc at any particular narrow brightness range, I tend to let skin ride somewhere between 45IRE and 70IRE depending on the overall scene.
If I can fit the contrast range of the scene into the 11.5 stops of a cinegamma then I will often use the cinegammas over S-Log because of the noise improvement. S-Log comes into it’s own where you have an extreme contrast range that needs to be captured. However at the end of the day you do still have to remember that the end display device is unlikely to be able to display more than 7 stops with any accuracy!
One tool I have found very useful is the BlackMagic HDLink box. I often use this to connect to a monitor as it has the ability to apply LUT’s very quickly. If you have a PC connected to the HDLink you can go in an modify the LUT curve in real time and in effect do an on-set grade. The HDLink is only $499 USD.
Well I posted here a few days ago about how Data was distributed across the S-Log curve. David williams (thanks David) questioned some of the things in my post raising some valid questions over it’s accuracy, so I withdrew the post in order to review it further. While the general principles within the post were correct (to the best of my knowledge and research) and I stand by them, some of the numbers given were not quite right and the data/exposure chart was not quite right.
Before going further lets consider the differences between the a video sensor works and the way our eyes work. A video sensor is a linear device while our own visual system is a logarithmic system. Imagine you are in a room with 8 light fittings, each one with the same power and light output. You start with one lamp on, then turn on another. When you turn on the second lamp the room does not appear to get twice as bright even though the amount of light in the room has actually doubled. Now with two lamps on what happens when you turn on a third? Well you wouldn’t actually notice much of a change. To see a significant change you would need to turn on 2 more lamps. Now with 4 lamps on to see a significant difference you would need to turn on a further 4 lamps. Only adding one or two would make little visual difference. This is because our visual system is essentially a logarithmic system.
Now lets think about F-Stops. An f stop (or T-stop) is a doubling or halving of exposure. So again this is a logarithmic system. If with one light bulb your scene is one stop then to increase the scene brightness by one stop you must double the amount of light, so you would add another light bulb. Now to increase the scene brightness by a further stop you would have to take your existing two light bulbs and double it again to 4 light bulbs, and so on… 2, 4, 8, 16, 32, 64….
Now going back to a video sensor, take a look at the illustrative graph below. The horizontal scale is the number of lightbulbs in our hypothetical room and the vertical scale is the video output from an imaginary video sensor in percent. Please note that I am trying to illustrate a point, the numbers etc are not accurate, I’m just trying to explain something that is perhaps miss-understood by many, simply because it is difficult to understand or poorly explained elsewhere. The important thing to note is that the plotted blue line is a straight line, not a curve because the sensor is a linear device.
Now look at this very similar chart. The only difference now is that I have added an f-stop scale to the horizontal axis. Remember that one f-stop is a doubling of the amount of light, not simply one more lightbulb. I have also changed the vertical scale to data bits. To keep things simple I’m going to use something close 10 bit recording which actually has 956 data bits or steps (bits 64 to 1019 out of 1024 bits), but lets just round that up to 1000 data bits to keep life simple for this example.
So we can see that this imaginary video sensor uses bits 0-50 for the first stop, 50-100 for the second stop, 100-200 for the third stop, 200-400 for the fourth and 400-800 for the fifth. So it is easy to see that huge amounts of data are required to record each stop of over exposure. The brighter the image the more data that is required. Clearly if you want to record a wide dynamic range using a linear system you need massive numbers of data bits for the highlights, while the all important mid tones and shadow areas have relatively little data allocated to them. This is obviously not a desirable situation with current data limited recording systems, you really want to have sufficient data allocated to your mid-tones so that in post production you can grade them satisfactorily.
Now look what happens if we allocate the same amount of data to each stop of exposure. The green line is what you get if, in our imaginary camera we use 200 data bits to record each of our 5 stops of dynamic range. Does the shape of this curve look familiar to anyone? The important note here is that compared to the sensors linear output (the blue line) as the image brightness increases less and less data is being used to record the highlights. This mimics the way we see the world and helps ensure that in the mid ranges where skin tones normally reside there is lots of data to play with in post. Our visual system is most acute in the mid range. that’s because some of the most important things that we see are natural tones, plants, fauna and people. We tend to pay much less attention to highlights as these are rarely of interest to us. Because of this we can afford to reduce the amount of information in video highlights without the end user really noticing. This technique is used by most video cameras when the knee kicks in and compresses highlights. It’s also used by extended gamma curves such as cinegamma’s and hypergamma’s.
Anyone that’s seen a hypergamma curve or cinegamma curve plot will have seen a similar shape of curve. Hypergammas and Cinegammas also use less and less data to record highlights (compared to a linear response) and in many ways achieve a similar improvement in the captured dynamic range.
Hypergammas are not the same as S-Log however. Hypergammas are designed to be useable without grading, even if it’s not ideal. Because of this they stay close to standard gammas in the mid range and it’s only really the highlights that are compressed, this also helps with grading if recording using only an 8 bit codec as the amount of pushing and pulling required to get a natural image is less extreme. However because the Hypergammas allocate more data in the 60 to 90 percent exposure range to stay close to standard gamma the highlights have to be more highly compressed than S-Log so there is less highlight data to work with than with S-Log. If we look at the plot below which now includes an approximate S-Log curve (pink line) you can see that log recording has a much larger difference from a standard gamma in the mid ranges, so heavy grading will be required to get a natural looking image.
Because of the amount of grading that will normally be done with S-Log, recording the output using a 10 bit recorder is all but essential.
When I wrote this article I spent a lot of time studying the Sony S-Log white paper and reading up on S-Log and gamma curves all over the place. One thing that I believe leads to some confusion is the way Sony presents the S-Log data curve in the document. The exposure is plotted against the data bits using stops as opposed to image brightness. This is a little confusing if you are used to seeing traditional plots of gamma curves like the ones I have presented above that plot output against percentage light input. It’s confusing as Sony forget that using stops as the horizontal scale means that the horizontal scale is a log scale and this makes the S-Log “curve” appear to be a near straight line.
I have not used S-Log on an F3 yet. It will be interesting to see how it compares to Hypergamma in the real world. I’m sure it will bring some advantages as it allows for an 800% exposure range. I welcome any comments or corrections to this article.
A lot of the issues with any camera and the dynamic range it can record are not due to limitations of the cameras hardware but to retain compatibility with existing display technologies, in particular the good old fashioned TV set that has been around for half a century. The issue being that in order for all TV owners to see a picture that looks “natural” there has to be a common standard for the signal sent to the TV’s that will work with all sets from the very oldest to the most recent.
As even the most recent TV’s and monitors often struggle to display a contrast range greater than 7 stops there is no point in attempting to feed them with more, Taking 12 stops and simply squashing it into 7 stops will create a disappointing, flat and dull looking image. So for productions where extensive grading is not taking place, it is not desirable to record information beyond that which the existing broadcast system can handle. This is why the vast majority of modern camcorders with the knee off and using a standard gamma curve all exhibit very similar dynamic ranges (7 to 8 stops typically), because the limitation is generally not that of the sensor, but that of the gamma curves used in broadcast television. By adding a bit of highlight compression through a cameras knee circuit we can stretch out the dynamic range a bit as our visual system is most acute to inaccuracies in the the mid ranges of an image where faces, people and natural subjects normally appear so we don’t tend to notice strong compression occurring in highlights such as the sky or reflections. A well designed knee circuit can help gain an extra 2 or 3 stops by compressing the hell out of highlights, but as most of us are probably aware it can create it’s own issues with the near complete loss of real detail in clouds and the sky as well as color saturation issues on skin highlights, this is gamma curve compression rearing it’s ugly head. Moving on, we come to cinegammas, hypergammas and other similar extended range gammas. One of the issues with a traditional aggressive knee circuit is that it is either on or off, compressing or not compressing, there is no middle ground and this makes grading problematic as it is all but impossible to extract any meaningful data from very highly compressed highlights. Cinegammas etc address this by slowly increasing the amount of compression used as image brightness increases. In addition the gamma curve compression starts much earlier, long before you get to what would traditionally be regarded as “highlights”. This slow and gentle onset of compression grades in a more pleasing manner than a conventional knee. If you don’t grade the added mid-to-highlight compression results in a picture that looks a little flat and lacks “punch”, but is not overly objectionable to view. There is however a limit to just how much data you can cram into a compressed codec or recording system. Cinegammas and Hypergammas are tailored to give optimum performance with existing 8 bit and 10 bit high compression systems and workflows so the design engineers chose to only record a range of about 11 stops as trying to extract more than this from systems essentially designed to only record 7 to 8 stops will lead to visible compression artefacts. Technologies have continued to advance and now it’s remarkably easy (compared to just a couple of years ago) to record 10 bits of 4:2:2 or 4:4:4 data without compression or with only minimal compression. By eliminating or at least significantly reducing the compression artefacts it’s now possible to extract more meaningful data from a compressed gamma curve than was possible previously. S-Log is in effect nothing more than a heavily modified gamma curve, taking cinegammas and hypergammas to the next level. S-Log needs 10 bit recording to work as the curve compression starts much lower in the curve, so when grading those crucial skin tones and natural objects will need to be un-compressed to look natural and 8 bits of data just would not give enough range. As the image brightness increases the amount of gamma curve compression is increased logarithmically. If you look at the data being recorded this means that the majority of the 10 bit data is allocated to shadow areas then mid tones with less and less data being used to record highlights.
Most modern cameras, not just the XDCAM’s simply ignore highlight information beyond what can be recorded, this results in the image getting clipped at a given point depending on the gamma curve being used. Interestingly using negative gain on a camcorder can act as a low end clip as very small brightness changes will be reduced by the negative gain, possibly to the point where they are no longer visible. This normally results in a reduction in dynamic range (as well as noise). I suspect this is why the F3 has less noise using standard gammas because the sensor has excess dynamic range for theses curves and good sensitivity, so Sony can afford to set the arbitrary 0db point in negative space without impacting the recorded DR but giving a low noise floor benefit. For S-Log however it’s possible to record a greater dynamic range so 0db is returned to true zero and as a result the noise floor increases a little.
LUT’s are just a reverse gamma curve applied to the S-Log curve to restore the curve to one that approximates a standard gamma, normally REC-709. They are there for convenience to provide an approximation of what the finished image might look like. However applying an off the shelf LUT will impact the dynamic range as an assumption has to be made as to which parts of the image to keep and which to discard as we are back to squeezing 12 bits into 7 bits. As every project, possibly every shot will have differing requirements you would need an infinite number of LUT’s to be able to simply hit an “add LUT” button to restore your footage to something sensible. Instead it is more usual for the colorist or grader to generate their own curves to apply to the footage. Most NLE’s already have the filters to do this, it’s simply a case of using a curves filter or gamma curve correction to generate your own curves that can be applied to your clips in lieu of a LUT.
The graph to the left shows and idealised, normal gamma curve for a video production chain. The main thing to observe is that the curve is in fact pretty close to a straight line (actual gammacurves are very gentle, slight curves). This is important as what that means is that when the filmed scene gets twice as bright the output shown on the display also appears twice as bright, so the image we see on the display looks natural and normal. This is the type of gamma curve that would often be referred to as a standard gamma and it is very much what you see is what you get. In reality there are small variations of these standard gammacurves designed to suit different television standards, but those slight variations only make a small difference to the final viewed image. Standard gammas are typically restricted to around a 7 stop exposure range. These days this limited range is not so much to do with the lattitude of the camera but by the inability of most monitors and TV display systems to accurately reproduce more than a 7 stop range and to ensure that all viewers whether they have 20 year old TV or an ultra modern display get a sensible looking picture. This means that we have a problem. Modern cameras can capture great brightness ranges, helping the video maker or cinematographer capture high contrast scenes, but simply taking a 12 stop scene and showing it on a 7 stop display isn’t going to work. This is where modified gammacurves come in to play.
The second graph here shows a modified type of gamma curve. This is similar to the hypergamma or cinegamma curves found on many professional camcorders. What does the graph tell us? Well first of all we can see that the range of brightness or lattitude is greater as the curve extends out towards a range of 10 T stops compared to the 7 stops the standard gamma offers. Each additional stop is a doubling of lattitude. This means that a camera set up with this type of gamma curve can capture a far greater contrast range, but it’s not quite as simple as that.
Un-natural image response area
Look at the area shaded red on the graph. This is the area where the cameras capture gamma curve deviates from the standard gamma curve used not just for image capture but also for image display. What this means is that the area of the image shaded in red will not look natural because where something in that part of the filmed scene gets 100% brighter it will only be displayed as getting 50% brighter for example. In practice what this means is that while you are capturing a greater brightness range you will also need to grade or correct this range somewhat in the post production process to make the image look natural. Generally scenes shot using hypergammas or cinegammas can look a little washed out or flat. Cinegammas and Hypergammas keep the important central exposure range nice an linear, so the region from black up to around 75% is much like a standard gamma curve, so faces, skin, flora and fauna tend to have a natural contrast range, it is only really highlights such as the sky that is getting compressed and we don’t tend to notice this much in the end picture. This is because our visual system is very good at discerning fine detail in shadow and mid tones but less accurate in highlights, so we tend not to find this high light compression objectionable.
S-Log Gamma Curve
Taking things a step further this even more extreme gamma curve is similar to Sony’s S-Log gamma curve. As you can see this deviates greatly from the standard gamma curve. Now the entire linear output of the sensor is sampled using a logarithmic scale. This allows more of the data to be allocated to the shadows and midtones where the eye is most sensitive. The end result is a huge improvement in the recorded dynamic range (greater than 12 stops) combined with less data being used for highlights and more being used where it counts. However, the image when viewed on a standard monitor with no correction that looks very washed out, lacks contrast and generally looks incredibly flat and uninteresting.
Red area indicates where image will not look natural with S-Log without LUT
In fact the uncorrected image is so flat and washed out that it can make judging the optimum exposure difficult and crews using S-Log will often use traditional light meters to set the exposure rather than a monitor or rely on zebras and known references such as grey cards. For on set monitoring with S-Log you need to apply a LUT (look Up Table) to the cameras output. A LUT is in effect a reverse gamma curve that cancels out the S-Log curve so that the image you see on the monitor is closer to a standard gamma image or your desired final pictures. The problem with this though is that the monitor is now no longer showing the full contrast range being captured and recorded so accurate exposure assessment can be tricky as you may want to bias your exposure range towards light or dark depending on how you will grade the final production. In addition because you absolutely must adjust the image in post production quite heavily to get an acceptable and pleasing image it is vital that the recording method is up to the job. Highly compressed 8 bit codecs are not good enough for S-Log. That’s why S-Log is normally recorded using 10 bit 4:4:4 with very low compression ratios. Any compression artefacts can become exaggerated when the image is manipulated and pushed and pulled in the grade to give a pleasing image. You could use 4:2:2 10 bit at a push, but the chroma sub sampling may lead to banding in highly saturated areas, really Hypergammas and Cinegammas are better suited to 4:2:2 and S-Log is best reserved for 4:4:4.
The new Sony PMW-F3 can record using normal standard gammas and hypergammas, which compress highlights and increase the dynamic range to 460% and S-Log which is, as it’s name suggests an “S” shaped logarithmic gamma curve that compresses both highlights and lowlights giving upto 800% dynamic range. However the noise figures given by Sony are very different for S-Log and Hypergammas.
It will be interesting to see what the ramifications of the F3?s, 63db noise figure using standard gammas and hypergammas, vs the 57db nf using S-Log.
S-Log on the F3 will give you almost an extra stop of dynamic range but doubles the amount of noise. In most grading situations noise is the defining factor as to how far you can push the image in post. With normal gammas, at 57db the noise level is at the point where 10 bit recording brings little advantage as the noise is still around the minimum sample size. But S-Log is different as the distribution of data across the gamma curve is not linear, you must have 10bit recording for S-Log to work correctly. With the standard gammas at 63db there will be a definite advantage to recording 10bit.
The noise behaviour suggests that the true zero point noise level for the F3 is 57db, but that for “normal” use, due to the sensors very large dynamic range I suspect that Sony have chosen to reduced the gain by 6db, thus reducing both the noise and dynamic range. In effect the camera is operating with -6db gain switched in as default. However the sensor is sensitive enough to still give excellent low light performance despite the reduced gain and has more than enough dynamic range to still give the 11 stops that can be recorded with Hypergammas. This points to excellent low light performance as with +6db of gain switched in (with standard or hypergammas) your still going to have a 57db noise figure and at +9db gain it should have about the same amount of noise as an EX1 at 0db! Ah.. the delight of big pixels.
Camera setup, reviews, tutorials and information for pro camcorder users from Alister Chapman.