Gavin Greenwalt
Well-known member
- Joined
- Dec 29, 2006
- Messages
- 7,614
- Reaction score
- 0
- Points
- 36
- Age
- 40
Stacey Spears and I set out to try and somewhat definitively determine what the dynamic range of Dragon is. To perform our test we used the DSC Labs XYLA-21 test chart. The first thing that should be noted about this chart is that it's a tungsten not daylight chart. In order to try and minimize as many variables and sources of error as possible I only used a single color channel to perform this test and debayered the sensor at RED's native daylight color temperature. One thing I discovered while testing was that depending on your white balancing the color science would sometimes 'borrow' sensitivity from one channel for another. This is to be expected since in the real world the debayer process is designed to leverage multiple colored photo sites in order to produce a single pixel, however it complicates and diminishes the accuracy of any dynamic range test that goes through RED's Color Science Black Box. Access to the RAW photosite data would have hugely simplified the process and increased accuracy. If I could change one thing in my testing procedures it would be to use a Dragon Monochrome where all color science questions would be eliminated.
It's important to keep in mind that this is also a very academic dynamic range test. We were not attempting to determine what the "usable" dynamic range of the sensor was as that is a subjective mark. Nor could we possibly cover any and every shooting situation and the myriad of variables that can affect dynamic range such as color temperature, gamma processing, colorscience, saturation etc.
For this chart I averaged a total of 72 frames (6 stops) in order to produce the final image. The effect of averaging 72 frames is that all of the random noise averages down quite a bit and you can see what is statistically information and what is just noise. This is not a real-world situation unless you were shooting stills. In a later post I would like to propose a possible way to more scientifically discuss "usable" dynamic range. I will point out though that after stop 14 even with 72 frames averaged there are still pixels that are black. I would wager that this is similar to DXO's evaluation in which even in a perfect world with infinite noise averaging -- beyond 14 stops you will still have noise. Another note is that similar to another thread about Dragon I did notice excessive red channel noise at times. In fact even within the frames I averaged there was more than a stop of variation from one second to the next. So while this is one test of dynamic range, the low end is the inconsistent end and just variability from one second to the next will affect precisely how many stops are detectable.
The other thing you will notice is that there is a vertical line on the left, this is because I stitched together two separate shots. While testing we noted that the lens we were using had a slight amount of vignetting. If we had framed up the chart and exposed for the top chip clipping the vignetting would have acted as a slight ND on the top chips and given us the illusion of extended dynamic range that did not exist. To that end we exposed the top chip in the center of the frame and then panned the camera to better center the dark frames in the center of frame to eliminate any vignetting. The pan also framed out the clipping top chips. This eliminated a bit of glare which could raise up a chip beyond black as well as eliminate what appears to be a bit of readout contamination in any CMOS sensor row that's clipping. Again, in a perfect world I would film each chip by itself so that there is no chance of contamination. And I would frame each chip centered in frame. However, I don't think it would be significant enough to affect the results.
It should be noted that for each pan I did a full exposure rack test using the iris to ensure that I did in fact hit exactly clipping. While I am 99% certain I hit the max exposure for both the shadow blend and the clipping blend it is feasible I was off by about 1/4 of a stop. However, I used a couple frames from a 1/4 stop under the target exposed for noise cleanup. It doesn't affect the results, it just makes the chart a little more clearer. Like I said, with 72 frames of blending this is by no means what you should expect to see in the real world. The bottom chips are literally 72x noisier on any given frame.
One source of error we weren't able to eliminate was whether or not the XYLA chart is in fact accurate. Based on the RED linear EXRs the output does appear to be pretty linear with each chart being 1/2 intensity of the previous however we did not have a light meter accurate enough to measure the darker stops.
REDcode was 4:1
Camera build was 5.2.15
Otus 55mm f/1.4 (Canon)
EXRs & Nuke Script: http://1drv.ms/1i21iRM
R3Ds: http://1drv.ms/1i21tMX
The most challenging part of this test and the part most likely to be wrong is that this is precisely the top chip is in fact "just barely under clipping". In linear space I can see the texture of the chip chart as I stopped down in 1/4 stop increments. As added reassurance when debayered linear space the intensity goes .99 -> 0.52 which is almost precisely one stop increment (stops are half intensity).
From my reading of the chart there are definitely 18 stops of dynamic range if you dig way down into the noise. The 18th -> 19th stop is a bit debateable. Since the XYLA test chart only has full stops there is no way to know precisely how much more than 18 there is.
Biggify: http://i111.photobucket.com/albums/...on_Dynamic_range_v10_zpsc2bc2d8e.png~original
It's important to keep in mind that this is also a very academic dynamic range test. We were not attempting to determine what the "usable" dynamic range of the sensor was as that is a subjective mark. Nor could we possibly cover any and every shooting situation and the myriad of variables that can affect dynamic range such as color temperature, gamma processing, colorscience, saturation etc.
For this chart I averaged a total of 72 frames (6 stops) in order to produce the final image. The effect of averaging 72 frames is that all of the random noise averages down quite a bit and you can see what is statistically information and what is just noise. This is not a real-world situation unless you were shooting stills. In a later post I would like to propose a possible way to more scientifically discuss "usable" dynamic range. I will point out though that after stop 14 even with 72 frames averaged there are still pixels that are black. I would wager that this is similar to DXO's evaluation in which even in a perfect world with infinite noise averaging -- beyond 14 stops you will still have noise. Another note is that similar to another thread about Dragon I did notice excessive red channel noise at times. In fact even within the frames I averaged there was more than a stop of variation from one second to the next. So while this is one test of dynamic range, the low end is the inconsistent end and just variability from one second to the next will affect precisely how many stops are detectable.
The other thing you will notice is that there is a vertical line on the left, this is because I stitched together two separate shots. While testing we noted that the lens we were using had a slight amount of vignetting. If we had framed up the chart and exposed for the top chip clipping the vignetting would have acted as a slight ND on the top chips and given us the illusion of extended dynamic range that did not exist. To that end we exposed the top chip in the center of the frame and then panned the camera to better center the dark frames in the center of frame to eliminate any vignetting. The pan also framed out the clipping top chips. This eliminated a bit of glare which could raise up a chip beyond black as well as eliminate what appears to be a bit of readout contamination in any CMOS sensor row that's clipping. Again, in a perfect world I would film each chip by itself so that there is no chance of contamination. And I would frame each chip centered in frame. However, I don't think it would be significant enough to affect the results.
It should be noted that for each pan I did a full exposure rack test using the iris to ensure that I did in fact hit exactly clipping. While I am 99% certain I hit the max exposure for both the shadow blend and the clipping blend it is feasible I was off by about 1/4 of a stop. However, I used a couple frames from a 1/4 stop under the target exposed for noise cleanup. It doesn't affect the results, it just makes the chart a little more clearer. Like I said, with 72 frames of blending this is by no means what you should expect to see in the real world. The bottom chips are literally 72x noisier on any given frame.
One source of error we weren't able to eliminate was whether or not the XYLA chart is in fact accurate. Based on the RED linear EXRs the output does appear to be pretty linear with each chart being 1/2 intensity of the previous however we did not have a light meter accurate enough to measure the darker stops.
REDcode was 4:1
Camera build was 5.2.15
Otus 55mm f/1.4 (Canon)
EXRs & Nuke Script: http://1drv.ms/1i21iRM
R3Ds: http://1drv.ms/1i21tMX
The most challenging part of this test and the part most likely to be wrong is that this is precisely the top chip is in fact "just barely under clipping". In linear space I can see the texture of the chip chart as I stopped down in 1/4 stop increments. As added reassurance when debayered linear space the intensity goes .99 -> 0.52 which is almost precisely one stop increment (stops are half intensity).
From my reading of the chart there are definitely 18 stops of dynamic range if you dig way down into the noise. The 18th -> 19th stop is a bit debateable. Since the XYLA test chart only has full stops there is no way to know precisely how much more than 18 there is.
Biggify: http://i111.photobucket.com/albums/...on_Dynamic_range_v10_zpsc2bc2d8e.png~original
