Also too often we can get over focused on one aspect of a camera when choosing to buy, i.e. I need a camera that can shoot in extreme low light because I have this shoot coming up and can only use a lamp as a light source or I need to shoot at a bar and can't bring in lights. Well camera A is very good in low light but lacking in some other areas and camera B is not quite as good in low light but has lots of other features that camera A doesn't. So you choose camera A but after you wrap up that gig you may not need that top tier low light performance but some of those other features on camera B would really be useful for future gigs. I'm mainly a wedding shooter and thought I needed the best low light camera out there so I went with the A7Sii and was happy with it for a while but there were times it just could not get the job done to my liking or caused unnecessary stress for one reason or another. I'd previously owned all the GH series going back to the GH1 so when the GH5 came out I was drawn back in with the features it offered and pairing it with the Voigtlanders offered close enough low light performance that I was confident I could take it into any environment and enjoy shooting with it.
GH5 How to get the best low light, high quality from a GH5
- Thread starter garya
- Start date
Chris, that's an interesting analysis.
But one thing that I think is missing is the relationship between the SNR of the sensor and the SNR of the video signal. They are not the same, because the sensor readout is downsampled to produce the video signal (usually, depending on the camera). Downsampling reduces noise.
Also, you mention a relationship between the number of bits in a video signal encoding and the maximum dynamic range. That relationship also depends on the color space encoding. Your numbers would seem to assume a linear 8-bit or 10-bit encoding. Video signals almost never use linear encodings, so I don't think those numbers apply to the types of 8-bit and 10-bit video signals that we usually talk about.
But one thing that I think is missing is the relationship between the SNR of the sensor and the SNR of the video signal. They are not the same, because the sensor readout is downsampled to produce the video signal (usually, depending on the camera). Downsampling reduces noise.
Also, you mention a relationship between the number of bits in a video signal encoding and the maximum dynamic range. That relationship also depends on the color space encoding. Your numbers would seem to assume a linear 8-bit or 10-bit encoding. Video signals almost never use linear encodings, so I don't think those numbers apply to the types of 8-bit and 10-bit video signals that we usually talk about.
cyvideo
Veteran
Balazer
Yes I would agree with you that down sampling does perceptually reduce the appearance of noise. You have no disagreement from me on that point. As to linear / non linear encoding the way I and many others in the industry would look at the encoding process differs somewhat with regards to your statement:
“Your numbers would seem to assume a linear 8-bit or 10-bit encoding. Video signals almost never use linear encodings...” Yes and no would be my response.
Agreed at a certain point in the processing chain. This is where the transform function of ‘Gamma’ is introduced but one must remember that the signal originating from the sensor is linear, goes through the transform and then has the non linear gamma correction removed to restore a linear light representation. At either end of the chain the image is linear. Linear at the start of the signal chain and linear at the end of the signal chain as perceived by the human eye.
I think a very concise description of the whole linear, non linear transfer function (gamma) back to linear is very concisely put by Charles Poynton who is probably considered amongst those of us who have worked in the areas of television engineering to be one of the leading experts in the field of video encoding. As he explains, if you wish to simulate the physical world, linear-light coding is necessary.
http://poynton.ca/~poynton/index.html
http://poynton.ca/~poynton/notes/TIDV/index.html
Quotes from Poynton’s ‘A Technical Introduction to Digital Video’:
“A video system should ideally satisfy the principle of superposition; in other words, it should exhibit linearity. A function f is linear if and only if (iff):
f(a+b) ≡ f(a)+f(b)
The function f can encompass an entire system: A system is linear iff the sum of the individual responses of the system to any two signals is identical to its response to the sum of the two. Linearity can pertain to steady-state response, or to the system's temporal response to a changing signal.
Linearity is a very important property in mathematics, in signal processing, and in video. But linearity in one domain cannot be carried across to another domain if a nonlinear function separates the two. An image signal usually originates in a sensor that has linear response to physical intensity. And video signals are usually processed through analogue circuits that have linear response to voltage or digital systems that are linear with respect to the arithmetic performed on the codewords. But a video camera applies a nonlinear transfer function - gamma correction - to the image signal. So the image signal is in a linear optical domain, and the video signal is in a linear electrical domain, but the two domains are not the same.
If you wish to simulate the physical world, linear-light coding is necessary. For example, if you want to produce a numerical simulation of a lens performing a Fourier transform, you should use linear coding. If you want to compare your model with the transformed image captured from a real lens by a video camera, you will have to “remove” the nonlinear gamma correction that was imposed by the camera, to convert the image data back into its linear-light representation.”
This is how I have understood the process of the linear / non linear / linear function of the digital video chain and to date it has served me well in the industry. Most engineers and techs I’ve worked with have also taken this to be the case. That it all starts linear at the sensor and ends up linear at the display. Signal processing is an interesting subject that one has to keep up with as it is constantly evolving. Love the discussion
Chris Young
Yes I would agree with you that down sampling does perceptually reduce the appearance of noise. You have no disagreement from me on that point. As to linear / non linear encoding the way I and many others in the industry would look at the encoding process differs somewhat with regards to your statement:
“Your numbers would seem to assume a linear 8-bit or 10-bit encoding. Video signals almost never use linear encodings...” Yes and no would be my response.
Agreed at a certain point in the processing chain. This is where the transform function of ‘Gamma’ is introduced but one must remember that the signal originating from the sensor is linear, goes through the transform and then has the non linear gamma correction removed to restore a linear light representation. At either end of the chain the image is linear. Linear at the start of the signal chain and linear at the end of the signal chain as perceived by the human eye.
I think a very concise description of the whole linear, non linear transfer function (gamma) back to linear is very concisely put by Charles Poynton who is probably considered amongst those of us who have worked in the areas of television engineering to be one of the leading experts in the field of video encoding. As he explains, if you wish to simulate the physical world, linear-light coding is necessary.
http://poynton.ca/~poynton/index.html
http://poynton.ca/~poynton/notes/TIDV/index.html
Quotes from Poynton’s ‘A Technical Introduction to Digital Video’:
“A video system should ideally satisfy the principle of superposition; in other words, it should exhibit linearity. A function f is linear if and only if (iff):
f(a+b) ≡ f(a)+f(b)
The function f can encompass an entire system: A system is linear iff the sum of the individual responses of the system to any two signals is identical to its response to the sum of the two. Linearity can pertain to steady-state response, or to the system's temporal response to a changing signal.
Linearity is a very important property in mathematics, in signal processing, and in video. But linearity in one domain cannot be carried across to another domain if a nonlinear function separates the two. An image signal usually originates in a sensor that has linear response to physical intensity. And video signals are usually processed through analogue circuits that have linear response to voltage or digital systems that are linear with respect to the arithmetic performed on the codewords. But a video camera applies a nonlinear transfer function - gamma correction - to the image signal. So the image signal is in a linear optical domain, and the video signal is in a linear electrical domain, but the two domains are not the same.
If you wish to simulate the physical world, linear-light coding is necessary. For example, if you want to produce a numerical simulation of a lens performing a Fourier transform, you should use linear coding. If you want to compare your model with the transformed image captured from a real lens by a video camera, you will have to “remove” the nonlinear gamma correction that was imposed by the camera, to convert the image data back into its linear-light representation.”
This is how I have understood the process of the linear / non linear / linear function of the digital video chain and to date it has served me well in the industry. Most engineers and techs I’ve worked with have also taken this to be the case. That it all starts linear at the sensor and ends up linear at the display. Signal processing is an interesting subject that one has to keep up with as it is constantly evolving. Love the discussion
Chris Young
...But what if the math and the actual real-world results are different.
Personally I would trust real-world side-by-side tests with both cameras shooting the same subject under identical conditions (not test charts, but the type of scenes one might encounter on a real shoot), and then judge the finished result from each camera. There is a good reason why some professional shooters have transitioned from the A7SII to the GH5, and it has nothing to do with math. They are getting better results with the GH5.
visceralpsyche
Well-known member
Shoot the A7S2 in Slog3 8bit 4:2:0, and the GH5 in V-LogL 10bit 4:2:2.
You'll quickly see the banding issues on the A7S2 when it comes time to convert both profiles back to a linear space, due to bit starvation because it only gets an effective 6.5 bits of color/tonal information to work with.
I agree that an 8bit linear profile using the entire width is sufficient for basic images without banding being obvious (until you grade heavily), but when it comes to log profiles which everyone wants to use for the added DR, it's not enough.
That's why the GH5 stands alone in this class currently.
I really, truly hope Sony gives us 10bit 4:2:2 and DCI 4K in the A7S3....
Cheers,
Paul
You'll quickly see the banding issues on the A7S2 when it comes time to convert both profiles back to a linear space, due to bit starvation because it only gets an effective 6.5 bits of color/tonal information to work with.
I agree that an 8bit linear profile using the entire width is sufficient for basic images without banding being obvious (until you grade heavily), but when it comes to log profiles which everyone wants to use for the added DR, it's not enough.
That's why the GH5 stands alone in this class currently.
I really, truly hope Sony gives us 10bit 4:2:2 and DCI 4K in the A7S3....
Cheers,
Paul
Great Blog link!
I would echo the point already made about DOF... if you can't keep your subject in focus, who cares? Nobody cares about the noisiness of an out-of-focus image. And even if you can make an ultra-thin DOF work for you in one type of situation, do you want to be restricted, creatively, to that one look every time you shoot in low light? Sometimes you want deep (or at least deeper) focus. And that's when you need the Sony, or to add light, or to switch your location.
The Sony is a big V8 muscle car, the GH5 is a Subaru WRX. If all you need to do is drag race, the muscle car is perfect, and it will beat the WRX every time at that one thing. If you need to go around some corners, or drive on snow or dirt, or go more than 100 miles on a tank, though...
The Sony is a big V8 muscle car, the GH5 is a Subaru WRX. If all you need to do is drag race, the muscle car is perfect, and it will beat the WRX every time at that one thing. If you need to go around some corners, or drive on snow or dirt, or go more than 100 miles on a tank, though...
...I think you've got the backwards. When you want deep or deeper focus ( i.e. more in focus ), that's when you need the GH5 because the Micro 4/3 sensor has a 2 F-stop DOF advantage at the same exposure.
So if you are shooting Full Frame with the A7SII at f/5.6 at 3200 ISO to get enough DOF to have things in focus, you can achieve the SAME DOF with the GH5 at f/2.8 at 800 ISO. So if you want deeper DOF ( i.e. more in focus ) then the GH5 is the camera you want. It's the compact high powered european sports car that leaves the V8 muscle car in the dust when it comes to deep DOF...
cyvideo
Veteran
Visceralpsyche and the Dingo (Canadian Dingo? Didn’t know you had any!)
"Shoot the A7S2 in Slog3 8bit 4:2:0, and the GH5 in V-LogL 10bit 4:2:2."
Agreed but this is the first time in this thread that LOG has been in contention. Grading of an image changes the perspective of the discussion somewhat when the original posters question was “What's the best settings and gear to get the highest quality 4k video with the the lowest available light?” What I was addressing was his reference to “... the lowest available light?” as this seemed to the major point of his question.
I would though also respectfully disagree with the point made that everybody want to shoote LOG. Although I fully accept your reasons as to why one would want to use LOG to try and maintain the quality of an image in post processing. Totally agree. Many do want to use LOG, mainly enthusiasts and professional, the number of which tends to be in the minority of general public users of cameras in the price range of the cameras we are talking about. That is why Panasonic marketed the LOG function as an optional purchase key. Panasonic maintain that this was made an option because the majority of GH buyers aren’t interested in the LOG upgrade.
Visceral states “I agree that an 8bit linear profile using the entire width is sufficient for basic images ...” I repeat, this is what I have been addressing here as I said my response was based on the OP’s point of “... the lowest available light?” No question as to whether he should shoot in 8 or 10-bit or LOG.
Again addressing the point of the OP’s point “...lowest available light?” All other factors such as shooting modes, frame rates, 8-bit or 10-bit, LOG or not being put aside I would draw your attention to a video whilst not done under strictly controlled circumstances demonstrates how one of the cameras, purely an 8-bit camera, can produce pretty quiet images at higher ISOs in low light. Now I don’t know whether he used the GH5 in 8-bit or 10-bit mode but the results from about 11:32 onwards shows how a well designed 8-bit process can perform. I know if I had to shoot at higher ISOs without the benefit of lighting, always light when I can, I would much rather try and extract the most usable images from the lower noise 8-bit images coming from, not an A7S, but an A6500 in this instance than the same ISO rated images from either the GH5 or the Canon C200 the other camera in this comparison. Any benefit to be gained in shooting 10-bit in this type of low light situation is totally negated by the fact that the image falls apart much earlier than the 8-bit image. Repeat, I don’t know if the GH5 was in 8 or 10-bit mode. If it was indeed in 8-bit mode then given the same encoding bit rate I would expect to see worse in 10-bit mode. Indeed images that have that much luma and chroma noise in them are hard to pull back regardless of how they were obtained be it in Rec 709 or LOG or any mode in between.
https://www.youtube.com/watch?time_continue=3&v=4rNtEEMIgrY
Dingo says “...But what if the math and the actual real-world results are different." In this real world example the clear winner in the comparisons of these three cameras with regards to noise in the “...lowest light possible?” is without a doubt the one that the math says should produce the lowest noise image. A lot of variables at play here though. The GH5 and the A6500 down sample so this helps reduce noise. The bigger sensor with its better signal to noise ratio helps the A6500. The whole signal processing chain and the encoding process in the Sony is visually performing better and quieter than either the GH5 or the C200 under this particular low light scenario. So yes I agree it’s not all to do with the pure math as there are many other design and processing factors involved that influence the final image. As I indicated earlier though give me the quietest image to work with whether it was shot Rec 709, LOG, RAW or on a piece of string. Starting out with the quietest image is most likely going to give one the best results regardless of whether it was obtained via an 8 or 10-bit process.
Love the discussion
Chris Young
"Shoot the A7S2 in Slog3 8bit 4:2:0, and the GH5 in V-LogL 10bit 4:2:2."
Agreed but this is the first time in this thread that LOG has been in contention. Grading of an image changes the perspective of the discussion somewhat when the original posters question was “What's the best settings and gear to get the highest quality 4k video with the the lowest available light?” What I was addressing was his reference to “... the lowest available light?” as this seemed to the major point of his question.
I would though also respectfully disagree with the point made that everybody want to shoote LOG. Although I fully accept your reasons as to why one would want to use LOG to try and maintain the quality of an image in post processing. Totally agree. Many do want to use LOG, mainly enthusiasts and professional, the number of which tends to be in the minority of general public users of cameras in the price range of the cameras we are talking about. That is why Panasonic marketed the LOG function as an optional purchase key. Panasonic maintain that this was made an option because the majority of GH buyers aren’t interested in the LOG upgrade.
Visceral states “I agree that an 8bit linear profile using the entire width is sufficient for basic images ...” I repeat, this is what I have been addressing here as I said my response was based on the OP’s point of “... the lowest available light?” No question as to whether he should shoot in 8 or 10-bit or LOG.
Again addressing the point of the OP’s point “...lowest available light?” All other factors such as shooting modes, frame rates, 8-bit or 10-bit, LOG or not being put aside I would draw your attention to a video whilst not done under strictly controlled circumstances demonstrates how one of the cameras, purely an 8-bit camera, can produce pretty quiet images at higher ISOs in low light. Now I don’t know whether he used the GH5 in 8-bit or 10-bit mode but the results from about 11:32 onwards shows how a well designed 8-bit process can perform. I know if I had to shoot at higher ISOs without the benefit of lighting, always light when I can, I would much rather try and extract the most usable images from the lower noise 8-bit images coming from, not an A7S, but an A6500 in this instance than the same ISO rated images from either the GH5 or the Canon C200 the other camera in this comparison. Any benefit to be gained in shooting 10-bit in this type of low light situation is totally negated by the fact that the image falls apart much earlier than the 8-bit image. Repeat, I don’t know if the GH5 was in 8 or 10-bit mode. If it was indeed in 8-bit mode then given the same encoding bit rate I would expect to see worse in 10-bit mode. Indeed images that have that much luma and chroma noise in them are hard to pull back regardless of how they were obtained be it in Rec 709 or LOG or any mode in between.
https://www.youtube.com/watch?time_continue=3&v=4rNtEEMIgrY
Dingo says “...But what if the math and the actual real-world results are different." In this real world example the clear winner in the comparisons of these three cameras with regards to noise in the “...lowest light possible?” is without a doubt the one that the math says should produce the lowest noise image. A lot of variables at play here though. The GH5 and the A6500 down sample so this helps reduce noise. The bigger sensor with its better signal to noise ratio helps the A6500. The whole signal processing chain and the encoding process in the Sony is visually performing better and quieter than either the GH5 or the C200 under this particular low light scenario. So yes I agree it’s not all to do with the pure math as there are many other design and processing factors involved that influence the final image. As I indicated earlier though give me the quietest image to work with whether it was shot Rec 709, LOG, RAW or on a piece of string. Starting out with the quietest image is most likely going to give one the best results regardless of whether it was obtained via an 8 or 10-bit process.
Love the discussion
Chris Young
...I think you've got the backwards. When you want deep or deeper focus ( i.e. more in focus ), that's when you need the GH5 because the Micro 4/3 sensor has a 2 F-stop DOF advantage at the same exposure.
So if you are shooting Full Frame with the A7SII at f/5.6 at 3200 ISO to get enough DOF to have things in focus, you can achieve the SAME DOF with the GH5 at f/2.8 at 800 ISO. So if you want deeper DOF ( i.e. more in focus ) then the GH5 is the camera you want. It's the compact high powered european sports car that leaves the V8 muscle car in the dust when it comes to deep DOF...
Yea, that's all true, except that in my (admittedly somewhat limited) experience, the Sony's ISO performance easily overcomes this difference. I've seen ISO100K footage that looked the same or cleaner than most of the ISO3200 footage I've seen from the GH5.
Shoot the A7S2 in Slog3 8bit 4:2:0, and the GH5 in V-LogL 10bit 4:2:2.
You'll quickly see the banding issues on the A7S2 when it comes time to convert both profiles back to a linear space, due to bit starvation because it only gets an effective 6.5 bits of color/tonal information to work with.
I agree that an 8bit linear profile using the entire width is sufficient for basic images without banding being obvious (until you grade heavily), but when it comes to log profiles which everyone wants to use for the added DR, it's not enough.
That's why the GH5 stands alone in this class currently.
I really, truly hope Sony gives us 10bit 4:2:2 and DCI 4K in the A7S3....
Cheers,
Paul
Using an external recorder like I suggested to the OP would 'level the playing field'. The 4.2.0 100mbps internal codec is totally inadequate for LOG capture but the 4.2.2 HDMI out in prores is another thing entirely.
visceralpsyche
Well-known member
No it won't, because it's mostly about 10 bit vs 8 bit (which you're not taking account of). 8 bit 4:2:2 output is still 8 bit, regardless of whether or not you record that 8 bit signal in 10 bit ProRes down the line.
Cheers,
Paul
No it won't, because it's mostly about 10 bit vs 8 bit (which you're not taking account of). 8 bit 4:2:2 output is still 8 bit, regardless of whether or not you record that 8 bit signal in 10 bit ProRes down the line.
Cheers,
Paul
Most of the artefacts ( particularly banding) are a product of high compression and not bit depth. 8bit 4.2.2 from the A7s HDMI in prores is very robust and won't break apart unless it's seriously abused.
Last edited:
visceralpsyche
Well-known member
8 bit Slog3 is roughly 6.5 actual bits of information usage. It's just not enough to prevent significant banding the moment you begin grading the picture (same issue as V-LogL as they're very close in terms of curves etc).
Compression can't compress what isn't there sadly.
Cheers,
Paul
Compression can't compress what isn't there sadly.
Cheers,
Paul
With the GH5 it's all about how things are handled in post. GH5 high ISO footage requires calibrated noise-reduction to eliminate most of the noise without harming the image. This means using something like NeatVideo and then building a NR profile for the settings that you want to shoot at, and then applying this profile to your footage once it's shot. GH5 ISO 3200 and sometimes ISO 6400 footage becomes very usable with proper NR in post.
The A7SII definitely has the low light advantage, but a lot will depend on what you shoot and how you shoot, to determine if you actually need an A7SII for the project.
But what if most of us don't need to shoot under "...lowest light possible". What if a GH5 at ISO 3200 with a f/1.2 lens and a dedicated NR profile in post covers 99 percent of all the low light work we will ever shoot ?
I can only talk about hands on experience with S-log 2 to an ext recorder. It takes huge tonal 'bending' to break it. Sure 10bit can be manipulated more but for sensible grades from properly exposed and WB'd footage it's solid and clean despite it's 'theoretical ' inferiority to 10bit. The GH5 is a great camera with a great internal codec but a full frame A7s ÷ recorder is also a serious capture device with significant advantages in low light work.8 bit Slog3 is roughly 6.5 actual bits of information usage. It's just not enough to prevent significant banding the moment you begin grading the picture (same issue as V-LogL as they're very close in terms of curves etc).
Compression can't compress what isn't there sadly.
Cheers,
Paul
Well then we'd be out of things to complain about. Can't have that!
visceralpsyche
Well-known member
Yes, Slog2 is basically using the full 8 bits, so it'll take more to break it than Slog3 (which is equivalent to the GH5's V-LogL).
Cheers,
Paul