John Brawley
Well-known member
.
Last edited:
Want high res without the compression ?
- Thread starter John Brawley
- Start date
combatentropy
Veteran
9.3K, 14 stops, global shutter. I wonder who makes the sensor.
ahalpert
Major Contributor
God.
Last edited:
John Brawley
Well-known member
Seems to be this one.
https://www.gpixel.com/products/area-scan-en/gmax3265-3-2-μm-65-mp-global-shutter-image-sensor/
Which is weird because it’s only 10bit at 70fps but they’re claiming 14-16 stops...
JB
https://www.gpixel.com/products/area-scan-en/gmax3265-3-2-μm-65-mp-global-shutter-image-sensor/
Which is weird because it’s only 10bit at 70fps but they’re claiming 14-16 stops...
JB
combatentropy
Veteran
That seems to match specs on everything but dynamic range. It says 66 dB, which is 11 stops.
---
10 bits should be enough for up to 19 stops, if they record log.
Bit depth doesn't really constrain the dynamic range of the camera. It sort of constrains the dynamic range of the monitor.
A bit is a binary digit and has 2 possible values, so 8-bit would be 2^8 (256), 10-bit is 2^10 (1,024), 12-bit is 2^12 (4,096), and so on. So in a 10-bit recording, there are 1,024 possible shades per channel, per pixel.
What are we trying to avoid? Banding. The consequence of a bit-starved image is banding. It is completely up to the human eye where banding begins, just like with resolution. The human eye starts to see banding when the shades per f-stop is less 70, but it depends on which f-stop we're talking about, because the dark areas of an image can get away with fewer shades (see Human vision and tonal levels, by Norman Koren).
But we're talking about when the human is watching the screen, not when the camera is watching the scene. So you would have to know how many f-stops are being output by the audience members' screens. This varies, of course, but it is probably 6-7 for the average middle-class viewer and 9-10 for home theater set-ups with HDR (movie theater screens also are about 10).
So dividing 1,024 by 10, you get 100 shades per f-stop, which should be enough to avoid the appearance of banding.
---
It's still more complicated than this, isn't it? Because the shades are evenly distributed only in a perfectly logarithmic recording. And the gamma curve of TVs and projectors isn't log. It's either Rec. 709 or whatever the latest standard is. It distributes it unevenly. But it's still nothing to worry about, because it is what makes 8-bit JPEGs look decent. You normally see banding in JPEGs only if someone has dropped the ball in the intermediate processing.
---
But it's still more complicated than this, isn't it? Because chances are you're going to push around the tones in post. This takes bands that were in one f-stop and pulls them out, leaving fewer in that f-stop. Hope there are still 70. It will push this band into a neighboring f-stop, which now has plenty of bands. So all this pushing and pulling in post really throws a wrench in things.
---
In general, though, 10-bit log is plenty. However you push and pull the image in post, as long as the final product has enough shades per f-stop (10-70, depending on which f-stop) and there are only 6-10 stops in the final product, then everything should be fine.
---
10 bits should be enough for up to 19 stops, if they record log.
Bit depth doesn't really constrain the dynamic range of the camera. It sort of constrains the dynamic range of the monitor.
A bit is a binary digit and has 2 possible values, so 8-bit would be 2^8 (256), 10-bit is 2^10 (1,024), 12-bit is 2^12 (4,096), and so on. So in a 10-bit recording, there are 1,024 possible shades per channel, per pixel.
What are we trying to avoid? Banding. The consequence of a bit-starved image is banding. It is completely up to the human eye where banding begins, just like with resolution. The human eye starts to see banding when the shades per f-stop is less 70, but it depends on which f-stop we're talking about, because the dark areas of an image can get away with fewer shades (see Human vision and tonal levels, by Norman Koren).
But we're talking about when the human is watching the screen, not when the camera is watching the scene. So you would have to know how many f-stops are being output by the audience members' screens. This varies, of course, but it is probably 6-7 for the average middle-class viewer and 9-10 for home theater set-ups with HDR (movie theater screens also are about 10).
So dividing 1,024 by 10, you get 100 shades per f-stop, which should be enough to avoid the appearance of banding.
---
It's still more complicated than this, isn't it? Because the shades are evenly distributed only in a perfectly logarithmic recording. And the gamma curve of TVs and projectors isn't log. It's either Rec. 709 or whatever the latest standard is. It distributes it unevenly. But it's still nothing to worry about, because it is what makes 8-bit JPEGs look decent. You normally see banding in JPEGs only if someone has dropped the ball in the intermediate processing.
---
But it's still more complicated than this, isn't it? Because chances are you're going to push around the tones in post. This takes bands that were in one f-stop and pulls them out, leaving fewer in that f-stop. Hope there are still 70. It will push this band into a neighboring f-stop, which now has plenty of bands. So all this pushing and pulling in post really throws a wrench in things.
---
In general, though, 10-bit log is plenty. However you push and pull the image in post, as long as the final product has enough shades per f-stop (10-70, depending on which f-stop) and there are only 6-10 stops in the final product, then everything should be fine.
ahalpert
Major Contributor
I believe you are incorrect, combatentropy, because your encoding bit-depth and your analog-digital conversion bit-depth are distinct.
The sensor signal correlates linearly with light intensity. Thus, with a 10-bit ADC, the numerical values that correspond to each stop of exposure are:
1st stop: 513-1024
2nd stop: 257-512
3rd stop: 129-256
4th stop: 65-128
5th stop: 33-64
6th stop: 17-32
7th stop: 9-16
8th stop: 5-8
9th stop: 3-4
10th stop: 1-2
If you do a dual gain output or use dithering, you can squeeze additional information out of the sensor.
But basically, if the dynamic range of the sensor surpasses the number of stops that can be distinguished at a certain bit-depth in a linear distribution, then each additional bit of accuracy in your ADC will yield an additional stop of dynamic range in the image.
The sensor signal correlates linearly with light intensity. Thus, with a 10-bit ADC, the numerical values that correspond to each stop of exposure are:
1st stop: 513-1024
2nd stop: 257-512
3rd stop: 129-256
4th stop: 65-128
5th stop: 33-64
6th stop: 17-32
7th stop: 9-16
8th stop: 5-8
9th stop: 3-4
10th stop: 1-2
If you do a dual gain output or use dithering, you can squeeze additional information out of the sensor.
But basically, if the dynamic range of the sensor surpasses the number of stops that can be distinguished at a certain bit-depth in a linear distribution, then each additional bit of accuracy in your ADC will yield an additional stop of dynamic range in the image.
morgan_moore
Major Contributor
So dividing 1,024 by 10, you get 100 shades per f-stop, which should be enough to avoid the appearance of banding.
100 to deliver 70 doesnt seem like enough to me, especially if you spread 1024 over 13 stops, 12bit seems the minimum to deliver real image quality.
100 to deliver 70 doesnt seem like enough to me, especially if you spread 1024 over 13 stops, 12bit seems the minimum to deliver real image quality.
combatentropy
Veteran
Ah you have a point. I believe in 10 bit so long as it is log. 10-bit linear will bit-starve the darker areas. So you can have either a logarithmic 10-bit ADC or use a linear ADC with 12-bit or higher and convert it to 10-bit log, preferably in camera.
You would be spreading 1024 over 13 only if the viewer's monitor was capable of 13 stops, which I think is unlikely. That would be 3 stops brighter than an Academy theater screen standard, and 3 stops is actually 16 times brighter (2^3).
Once it is captured, you need to forget about however many stops were in real life. The bits do not know.
I could make a 10-bit frame of a 12-stop scene, or a 10-bit frame from a 7-stop scene, or I could take a digital pen and draw a 10-bit image. No matter what, they're all going to be displayed on someone's 7-stop monitor (roughly).