KOMODO....

[Misha Engel]

No, that's not dynamic range, no it doesn't get increased whatsoever by scaling down and it is still nonsense.

Can you explain what is incorrect (or "nonsense" in your words) with the presented explanation in my thread?

 

[Phil Holland]

Can you explain what is incorrect (or "nonsense" in your words) with the presented explanation in my thread?

Okay.

What happens by downscaling and how is the dynamic range increased?

Well, the signal value of four neighboring pixels has a high degree of correlation (luminance will be roughly the same), whereas noise for those pixels in general should show no correlation as it is random.

Now, by downscaling (or averaging) four pixels of an 3840×2160 image to 1 pixel of an 1920×1080 image the signal value should stay very much the same, but the noise value will reduce because noise is random (uncorrelated), hence averaging (or in mathematical terms the root mean square value of noise which is measured in the recorded image) scales with the inverse square root of the number of samples that we average.

Hence 4 pixels into 1 gives a factor of two (sqrt(4)=2) higher signal to noise (SNR) ratio. Higher signal to noise ratio for a given luminance = higher dynamic range.

But you sort of state it in your explanation as to why it's incorrect.

Downsampling will decrease apparent image noise. You are correct about that.

In terms of Total Captured Dynamic Range or however you choose to measure DR that is a fixed measurement.

Depending on your tolerance for image noise you may feel you can use more of that "noisy DR down there" because it might less noisy, but that's where the magic stops.

In no way, shape, or form are you effecting how much Dynamic Range you have captured by down sampling.

 

[Björn Benckert]

To me what Misha says is true if you add ”usable” infront of DR. What ever the camera can capture that is not really usable is possibly interesting for some but for most people that work with images I think the latitude where the image still looks good is whats interesting. As most people use their cameras to try capture beautiful pictures, not grascale strips of measurable DR.

And as signal to noise ratio decrease when downsapling, that very much has an effect on what people consider is usable DR. At lower resolution it might look just fine to bump up the shadows by a few stops, for the same image that would not look good when doing so before downsampling and looked at in a higher resolution.

So I dont think Misha means to say that downsampling would bring in information from outside of the cameras measurable DR. To me he explains that quite well.

 

[Misha Engel]

Okay.




But you sort of state it in your explanation as to why it's incorrect.

Downsampling will decrease apparent image noise. You are correct about that.

In terms of Total Captured Dynamic Range or however you choose to measure DR that is a fixed measurement.

Depending on your tolerance for image noise you may feel you can use more of that "noisy DR down there" because it might less noisy, but that's where the magic stops.

In no way, shape, or form are you effecting how much Dynamic Range you have captured by down sampling.

DR is not a fixed measurement and depends on the S/N ratio, when you downscale a UHD/4k-DCI rgb frame to a FHD/2k-DCI rgb frame you gain one stop of DR in the shadows, the signal stays the same the noise is reduced by a factor of (sqrt4) 2, hence 1 more stop of DR.

Noise tolerance is no issue when you always measure in the same way under the same conditions (I like to stick to a SNR=2, and get billable stops, like ARRI does).

 

[Phil Holland]

How about now Bjorn?

 

[Björn Benckert]

Yes, now it gets more unclear.

What misha? You are saying that if you shoot an image and there is stuff not registerd due to low exposure... you are saying that you can downscale the image and all of a sudden you see things in the dark?

Thats not how it is. As I see it, there is a floor and below it the sensor cant see a thing. That is what is considered the bottom end of the measurable DR. Scaling down your image will not enhance that value.

Scaling down your image can only enhance the image captured within the measurable DR. In other words downsampling will increase the usable DR.

 

[Rakesh Malik]

Scaling down your image can only enhance the image captured within the measurable DR. In other words downsampling will increase the usable DR.

Ergo, the dynamic range of a given camera is fixed, unless you change the sensor.

 

[Björn Benckert]

Ergo, the dynamic range of a given camera is fixed, unless you change the sensor.

Correct, but the usable DR is as I see it relative to mastering resolution. As noise ratios comes in play.

Usable DR is quite different if using monstro to shoot for HD mastering or using it for a large highress print.

 

[Rakesh Malik]

Correct, but the usable DR is as I see it relative to mastering resolution. As noise ratios comes in play.

You can clean the shadows with noise reduction, and downsampling is one way to implement noise reduction -- a good one, obviously, but the recorded dynamic range remains constant.

Usable DR is quite different if using monstro to shoot for HD mastering or using it for a large highress print.

True, but that's because there's a subjective component in determining how much shadow noise is OK.

 

[Blair S. Paulsen]

What if the claim for the impact of downsampling was specifically about S/N ratio instead of DR?

To me, DR is the spread from D-min to D-max. Darkest pixel above absolute black to the brightest unclipped pixel.

S/N ratio is dependent on how noise is measured - perceptual, difference mattes, onset of "manufactured pixels", etc. That said, once a common metric is established it does allow mathematical comparisons.

Usable or, "paycheck" stops, is generally from just above the noise floor to clip. As a quick rule of thumb, I subtract the bottom 2 stops from whatever is claimed. If I really need every stop I can get for a particular shot/scene it's test, test, test. As a parting note, the ability of a device to accurately track both hue and saturation values 3 stops or more under middle gray typically improves the perceived S/N ratio.

Cheers - #19

 

[Aaron Lochert]

I think the argument is that if the bottom of the image is mucked up by noise, is that an accurate measurement of the detail in that area of the image? The 4-to-1 downscale provides more samples of that area of the image, increases the signal to noise ratio, thus getting a more accurate measurement there. But it does not add anything that wasn't already captured.

It's semantics to argue total DR and usable DR. And that's what Imatest is trying to do with their software, define what are the accurately measured stops and what are the muddy stops.

All that said, Imatest is only measuring a single frame. What we work with is temporal noise -- so we have an easier time as viewers seeing through the noise to the actual signal under it because the noise changes sporadically frame to frame and the detail under it either changes predictably in motion or doesn't change because it's a static frame. But either way, our brains will sort of average out the noise as we watch images in motion. So as DP's I think what we care more about is closer to the total DR as captured by the camera system and not so much the usable DR as measured for a still photographer.

 

[Misha Engel]

Yes, now it gets more unclear.

What misha? You are saying that if you shoot an image and there is stuff not registerd due to low exposure... you are saying that you can downscale the image and all of a sudden you see things in the dark?

Thats not how it is. As I see it, there is a floor and below it the sensor cant see a thing. That is what is considered the bottom end of the measurable DR. Scaling down your image will not enhance that value.

Scaling down your image can only enhance the image captured within the measurable DR. In other words downsampling will increase the usable DR.

Don't try to explain what you already explained.

I'll stick to my first explanation.

What happens by downscaling and how is the dynamic range increased?

Well, the signal value of four neighboring pixels has a high degree of correlation (luminance will be roughly the same), whereas noise for those pixels in general should show no correlation as it is random.

Now, by downscaling (or averaging) four pixels of an 3840×2160 image to 1 pixel of an 1920×1080 image the signal value should stay very much the same, but the noise value will reduce because noise is random (uncorrelated), hence averaging (or in mathematical terms the root mean square value of noise which is measured in the recorded image) scales with the inverse square root of the number of samples that we average.

Hence 4 pixels into 1 gives a factor of two (sqrt(4)=2) higher signal to noise (SNR) ratio. Higher signal to noise ratio for a given luminance = higher dynamic range.

 

[joshua csehak]

Instead of arguing academics, why not just show us an example?

 

[Sergio Perez]

What I want to know is... Any News update? Please?

 

[Misha Engel]

Instead of arguing academics, why not just show us an example?

I'll give it a shot (still a bit academic).

Assume you have a 4k sensor with a full well capacity of 40,000 e- and a read noise of 1.5 e-.
The S/N-ratio is 40,000/1.5=26,667 or 10log26667/10log2 = 14.7 stops or 6.0206 x 14.7 stops = 88.5 dB

When you don't need a 4k output but a 2k output, you can combine 4 pixels.
The S/N-ratio is now 4x26,667/(sqrt(4x1)) = 53,333 or 10log53,333/10log2 = 15.7 stops or 6.0206 x 15.7 stops = 94.5 dB

The sensors DR doesn't change but the resulting output has a higher DR at the cost of a lower resolution.

The noise value will reduce because noise is random (uncorrelated), hence averaging (or in mathematical terms the root mean square value of noise which is measured in the recorded image) scales with the inverse square root of the number of samples that we average.

Hence 4 pixels into 1 gives a factor of two (sqrt(4)=2) higher signal to noise (SNR) ratio. Higher signal to noise ratio for a given luminance = higher dynamic range.

AFAIK With TNR you look at noise from neighbouring frames and try to determine where the noise is and what the signal is, you can still apply this to the downscaled image.


For us this means the we can get the same DR with a FHD/2K-rgb output with our UMP's(full sensor readout) as you can with an ARRI AMIRA(full sensor readout).

 

[Karim D. Ghantous]

In no way, shape, or form are you effecting how much Dynamic Range you have captured by down sampling.

To the best of my knowledge, on a per-pixel level, DR is fixed and immutable at a fundamental level. Is that something that everyone can agree upon?

Oversampling, AFAIK, does increase usable DR, but oversampling exists so that downsampling can work. If you want to ride a bike really fast down a hill, you have to walk it up to the top beforehand. ;-) Maybe this is why we're all having semantic arguments on this point.

I would bet some money that a sheet of 4x5" film would have more usable latitude than a 24x36mm frame of that same emulsion, assuming the same AOV for each frame. I have no empirical evidence for this, but it just stands to reason.

What we work with is temporal noise -- so we have an easier time as viewers seeing through the noise to the actual signal under it because the noise changes sporadically frame to frame and the detail under it either changes predictably in motion or doesn't change because it's a static frame. But either way, our brains will sort of average out the noise as we watch images in motion. So as DP's I think what we care more about is closer to the total DR as captured by the camera system and not so much the usable DR as measured for a still photographer.

A good point to raise. I recall a thread where it was suggested that you could squeeze more usable DR out of any sensor by shooting at double your intended frame rate, and averaging pairs of frames.

The sensors DR doesn't change but the resulting output has a higher DR at the cost of a lower resolution.

That makes sense to me, whatever the engineers would say about it. Like I suggested above, downsampling works only because you oversample first.

 

[Hrvoje Simic]

I'll give it a shot (still a bit academic).

Assume you have a 4k sensor with a full well capacity of 40,000 e- and a read noise of 1.5 e-.
The S/N-ratio is 40,000/1.5=26,667 or 10log26667/10log2 = 14.7 stops or 6.0206 x 14.7 stops = 88.5 dB

When you don't need a 4k output but a 2k output, you can combine 4 pixels.
The S/N-ratio is now 4x26,667/(sqrt(4x1)) = 53,333 or 10log53,333/10log2 = 15.7 stops or 6.0206 x 15.7 stops = 94.5 dB

The sensors DR doesn't change but the resulting output has a higher DR at the cost of a lower resolution.

No it doesn't work that way.

If you reduce signal noise in shadow extremes with any of the techniques you don't gain any DR. What. So. Ever.
Poing of clipping and sensitivity stay the same. What is captured is captured. Period.

You only improve one image property of the range you are able to capture.
And there are more properties.

But some gear-"review" (promotion) sites might appreciate the continuance of the illusion.

The noise value will reduce because noise is random (uncorrelated), hence averaging (or in mathematical terms the root mean square value of noise which is measured in the recorded image) scales with the inverse square root of the number of samples that we average.

Hence 4 pixels into 1 gives a factor of two (sqrt(4)=2) higher signal to noise (SNR) ratio. Higher signal to noise ratio for a given luminance = higher dynamic range.

AFAIK With TNR you look at noise from neighbouring frames and try to determine where the noise is and what the signal is, you can still apply this to the downscaled image.

This is terminology issue, aided with some online SNR.

You are currently missing a distinction between:
- signal to noise ratio ("the relation of what is captured")
and
- dynamic range ("range of what gets captured")

That online statement you are using as a basis doesn't help in making that distinction.

For us this means the we can get the same DR with a FHD/2K-rgb output with our UMP's(full sensor readout) as you can with an ARRI AMIRA(full sensor readout).

Yes there is a name for this type of route to attributing meaning.

https://en.wikipedia.org/wiki/Wishful_thinking


That camera has potential and can shine in right hands, but wouldn't have that amount of stops if it was rigged to Orient Express.

Also, who is incompetent to capture a great image on that camera, no amount of additional stops avaiable in the extremes by the superior one will help to make the middle part of range appealing.

 

[Christoffer Glans]

Isn't what Misha is saying about the usability of the lowest range of the DR? Meaning, if the sensor has 14 stops, it still has 14 stops, but the lowest of those is unusable due to noise at the sensor resolution. By downsampling, the noise gets reduced, meaning you can see more clearly at that lowest stop of DR and therefore it is more usable than at top resolution. Feels like this is what is meant by the usable vs total DR definitions.

 

[joshua csehak]

I'll give it a shot (still a bit academic).

I'm sorry, by "example" I meant an actual movie clip where you downsample and we can see a gain in DR. B/C if we can't see it, it by definition isn't there, right?

 

[Christoffer Glans]

The myth that Bayer raw spatial color filter distribution produces chroma resolution equivalent to video codec subsampling dies hard, but it is definitely not true for any decent modern deBayer algorithm and uncompressed Bayer raw data. Raw compression schemes no doubt have some effect on that though.

Care to explain further? I only mean that you gain color sampling by downscaling, which is true for bayer sensors. As far as I know there are no Bayer sensors that work like a 3CCD in color sampling.

But you're not making money by possessing your camera; you're making money with the work you do, which could conceivably be just as good using any number of cameras. I think cameras are closer to expensive tools than investments. Investments should have the goal of increasing in value for profit, which digital cinema cameras as a rule don't do. Considering how quickly they become obsolete and then no longer hold value, they're closer to a computer (tool) than, say, real estate (investment).

Am I taking you too literally?

If we're talking about investments that generate the highest income in relation to the investment cost, there are far better cameras for that. Companies I've talked to that bought an URSA 4K said it was the best investment they've ever done. Their quality towards clients was beyond enough and after just two gigs they got the money back for the investment. Three years in, they have far surpassed the investment. With a Red, it wouldn't have been the same.

So yes, as an investment, there are far better options to invest in than Red. You go with Red if you need that quality, if you want that quality and if the project demand that quality. If you get what you need out of less expensive options, why would you choose a Red as an investment?