23 stops of range?

[Matthew Greene]

On a CMOS sensor you could control the gain of each pixel independently on capture. If this is real, that's what I would guess that they're doing in this sample. The image itself can't prove how many stops of range they're acomplishing without further data. Could be 10, could be 23.

 

[Dan Hudgins]

Threshold...

Threshold...

I understand how it worked in film: film has a nonlinear response, so flare (or even flashing) caused more of the grains into the higher response range where just a few photons would set them off. Digital, though, is linear, so even at the lowest levels (shadows) the response is the same.

Is it because the flare helps some of the shadow details to get further from the noise floor?

The A2D converter has a threshold. The pixels have a noise level. At low light levels the bias light introduces chaos.

If you modulate the bias light you can imporve the chaos.

In order to extract detail below the LSB the chaos needs to lift the pixel above the lower threshold of the A2D some of the time.

If you can record the chaos on a large enough sensor area and average the pixels you can get values smaller than the LSB of the A2D. If you can average many exposures at the same sensor area you can get values smaller than the LSB of the A2D.

Anything you do to reduce noise will defeat the gains. Did you shoot JPG rather than RAW? Newer cameras do not shoot true RAW data, so you might not record enough of the chaos.

Being linear or not is not the issue, it is the chaos being recorded since photons are integer to all sensors or film.

The larger sensors RED CO. is coming out with could give results of more than 14bits if you reduce the size of the images (with a resize that averages groups of pixels so all pixels are used, not a sub-sample) before compression, and control the chaos in the bias, i.e. white or pink noise in the bias source mixed in with the sensor signal analog before the A2D.

This works the same way dithering an 8bit image at high resolution gives you the tonal range of a 10bit image on the same monitor. If you use compression these tricks will not work because compression removes the low contrast detail.

If you average all the pixels in an image (or a gray card) and make a series of images with the light source brightness increasing very slowly, if all the pixels flip their binary values at the same time you would get stepped brightness changes, but if the binary values of the pixels flip in different frames then the average will climb on an almost continuous slope. If you only average part of the sensor you will see more noise in the slope than you will if you average the whole sensor.

Black and White film records only clear and opaque spots as grain, what looks like the continous tones of gray is the average of many on/off events at noisy thresholds.

 

[Wolf]

This question came up again so I'm bumping this thread.

 

[Wolf]

Not sure why so many otherwise smart people seems to think this is a hoax.

A little search around various cinematography forums reveal that several well respected DPs have been shown images from prototype sensors where Thomson Grass Valley used individual exposure control of each CMOS pixel to increase dynamic range. That was around 2007/2008... I don't know if this is the same thing, but that kind of dynamic range is nothing new from these guys.

Might also be a good idea to check the track record of Thomson Grass Valley (formerly Philips, BTS, Grass Valley...) before posting all sorts of conspiracy theories. These are the guys who made the Spirit 4K film scanner, Viper Filmstream camera, invented the Frame Transfer CCD to name a few...

Just saying.