5D Mark II Shutter Exposed! (Part II)

Mark Hahn · March 28th, 2009, 04:09 AM

(I apologize in advance for all these posts. I shouldn't be using the forum to "think out loud").

I found a good explanation (similar to my last theory) here ...

Point Grey Research (PGR) - Support - Knowledge Base

"The time delay between a row being reset and a row being read is the integration time. By varying the amount of time between when the reset sweeps past a row and when the readout of the row takes place, the integration time can be controlled. Since the integration process moves through the image over some length of time, some motion blur may become apparent".

Then he used a fan for an example (what a coincidence) ...

Note that the blade of the fan is stretched longer, not shorter. This makes the values in my original post start to make some sense. The faster number in each result is probably correct.

I need to figure out how to use my current images, or to change my setup, to correctly determine the real shutter speed to some accuracy, which was my original goal. I think I understand things well enough now to try that.

Jon Fairhurst · March 28th, 2009, 12:57 PM

Quote:

Originally Posted by Mark Hahn

Note that the blade of the fan is stretched longer, not shorter.

Certainly, one blade is long. Are you sure that the other two blades aren't compressed to be shorter as well?

I guess it depends on the direction of the sensor scan as well as the fan (moving upwards on one side and down on the other.)

One solution would be to film the fan from the top. As long as there's no vertical movement across the sensor, rolling shutter won't be an issue.

Mark Hahn · March 28th, 2009, 01:09 PM

Quote:

Originally Posted by Jon Fairhurst

Certainly, one blade is long. Are you sure that the other two blades aren't compressed to be shorter as well?

I guess it depends on the direction of the sensor scan as well as the fan (moving upwards on one side and down on the other.)

One solution would be to film the fan from the top. As long as there's no vertical movement across the sensor, rolling shutter won't be an issue.

I think I know how to measure exposure time very accurately with no fan, turntable, or any other moving object.

I understand now that at any one time the sensor has a band moving down it. The bottom of the band is the row being erased and the top of the band is the row being read. If the true speed was 1/60, the band would be half the height of the screen. 1/120 would be one quarter the height etc.

So here's the idea. I'm going to manually flash a fast strobe light which will allow me to see the band clearly. By measuring the height of the band (counting scan lines?) I will be able to precisely know the exposure time. I'll show results asap.

Mark Hahn · March 28th, 2009, 03:05 PM

OK, I've got some results from the flash experiment. That sure was easy. The setup was just the camera pointing at a wall and me flashing the strobe manually. I could see the band clearly on the LCD on every flash so I kept flashing until I caught one where the band was wholly contained on the screen. This turned out to be quite hard for slow speeds. At 1/40 it was usually off the top or bottom or wrapped around with some on the top and some on the bottom. I had to take dozens of flashes to get one in the screen.

Here are a few pics at 1/60, f2, ISO 100. The left band, which is only partially visible, is 362 scan lines (33% of the height) and the right band is 435 scan lines (40%). Because the left is shorter, we know there is idle time of at least 7% where there is no scan happening.

The height of the full band was less than 50%. The dead time would explain this because the rolling would have to be faster than 1/30 sec from top to bottom. If the exposure really was 1/60, then that would make the rolling rate 26,100 rows/sec (435*60). This is 38.3 usecs per row.

The flash time of my 430ex is rated at 1.2ms or less, which might be a problem. I set it for lowest power which I'm hoping would be the shortest time. We'll see if the numbers indicate a problem.

If this is all correct I can now determine the exact exposure time using this rolling rate and the number of lines exposed. Let's check this against some other speeds.

1/40 869 1/30.03
1/50 885 1/29.49
1/60 435 1/60 (duh)
1/80 436 1/59.8
1/125 338 1/77.2 (?)

The flash speed doesn't seem to be a problem because the 1/50 and 1/60 numbers were almost exactly 2 to 1.

But once again I have a problem. I am getting 1/30 readings which is impossible because I know there is dead time. This tells me 1/60 isn't really 1/60. This would also explain the weird 1/77.

So what is the real rolling rate? How can I figure it out if I can't trust any of the readings? I need to do some more thinking. Any ideas?

P.S. Jon and my results are in agreement that 1/40 and 1/50 are the same.

Mark Hahn · March 28th, 2009, 04:23 PM

I've cracked the code. On the 1/40 video (speed doesn't matter) I knew there were 869 lines in an exposure. I found a pair of adjacent frames where the bottom was exposed on the first and the top on the second. I added up the heights of each band and subtracted that from 869 and found the number of "lines" in the idle time. Adding this number to 1080 and multiplying by 30 gave me a rolling rate of 43,410 rows/sec.

I recalcuated the results I measured before using this rate. The numbers on the left are what I suspect are the actual values and the right side are the exact measurements. The difference from the actual numbers is negligable.

1/40 => 1/50 (869 => 1/49.95)
1/50 => 1/50 (885 => 1/49.05)
1/60 => 1/100 (435 => 1/99.79)
1/80 => 1/100 (436 => 1/99.56)
1/125 => 1/125 (338 => 1/128.4)

So It is simple. In this range it is either 1/50, 1/100, or 1/125. The 125 is a bit weird but I think it is real. I might measure some higher ones but I don't care. If anyone cares feel free to beg for me to shoot some more tests.

I think my work is done here. I believe these numbers are correct because they are so clean and logical. They are usable when designing any scheme for controlling that camera. I now know I can use either 1/40 or 1/50 and get 1/50, which not only gives repeatable results between scenes, but is very close to my holy grail of 1/48. Others who want faster speeds can use either 1/60 or 1/80 and always get 1/100.

Jon: I'm sorry but your holy grail of 1/60 is impossible.

Bernard Racelis · March 29th, 2009, 07:20 PM

Quote:

Originally Posted by Mark Hahn

If the true speed was 1/60, the band would be half the height of the screen. 1/120 would be one quarter the height etc.

If I'm reading your post correctly, it seems you're mixing up two different things.
The 'shutter' exposure time (the amount of time the pixels are charged) is not the same as the amount of time it takes to read/reset the lines.

For example, a pixel is charged/exposed for 1/2000th second, and that value is stored in the pixel. But it could take longer (let's say 1/60 second) to read/reset the values stored in all the pixels and lines in the whole frame.

If you're trying to measure the exposure time, you'll need to find a way to measure/compare the time the pixels are charged/exposed, not how fast the lines are read.

Mark Hahn · March 30th, 2009, 11:18 AM

Quote:

Originally Posted by Bernard Racelis

If I'm reading your post correctly, it seems you're mixing up two different things.
The 'shutter' exposure time (the amount of time the pixels are charged) is not the same as the amount of time it takes to read/reset the lines.

For example, a pixel is charged/exposed for 1/2000th second, and that value is stored in the pixel. But it could take longer (let's say 1/60 second) to read/reset the values stored in all the pixels and lines in the whole frame.

If you're trying to measure the exposure time, you'll need to find a way to measure/compare the time the pixels are charged/exposed, not how fast the lines are read.

If what you were saying was true, there wouldn't be a crisp band created by the flash. Also my numbers wouldn't have worked out.

How do you propose the exposure time is different than the rolling speed of the read/reset lines? There would have to be a global shutter like CCDs have.

Bernard Racelis · March 30th, 2009, 02:18 PM

Quote:

Originally Posted by Mark Hahn

If what you were saying was true, there wouldn't be a crisp band created by the flash. Also my numbers wouldn't have worked out.

Actually, what I was saying agrees with what you were showing -- the partial exposure (band) created by the flash. For example, if the flash duration is 1/250, and let's say it takes takes 1/60 to sweep the whole frame then only part of the frame will be exposed to the flash. And again, 1/250 (flash duration) and 1/60 (the time to sweep the whole frame) are separate from the shutter (which can be 1/125 for example).

CMOS cameras can go as high as 1/2000 'shutter'.
If the amount of time it takes to sweep the whole frame is proportional to the shutter/exposure which I think you were suggesting,
then if the shutter were 1/2000 and the whole frame can be read as fast as 1/2000 or so, then we wouldn't see as much skew / wobble.

Quote:

Originally Posted by Mark Hahn

How do you propose the exposure time is different than the rolling speed of the read/reset lines? There would have to be a global shutter like CCDs have.

If I, or anyone, have an easier way to test (other than the methods that have already been demonstrated -- to compare the motion blur and/or exposure ('brightness') of the video frame compared to a still shot), then we probably would have done it already ;)

Bernard Racelis · March 30th, 2009, 03:29 PM

Mark,

It could also be that I did not understand exactly the method that you were using to perform the test -- I hadn't bothered reading all the posts in detail since you've said that this thread has become cluttered. Perhaps you can explain in a simple/short way what exactly you're measuring, without all the clutter.

Jon Fairhurst · March 30th, 2009, 03:48 PM

Mark, Bernard,

I think the analysis is correct (and brilliant, BTW.) It's kind of the quantum physics of sensor analysis - it uses probability, in a way.

Assume that the flash of light is infinitely brief. At a given point of time, some sensors are held in reset (not being exposed), and others are being exposed at ambient (darker) levels. The exposure rolls from line to line. When the exposure is long, many lines are being exposed at the same time. When the exposure is short, few lines are being exposed simultaneously. Then BAM, the flash happens, and we get to see exactly how many lines were open simultaneously.

Mark shows the delay between lines to be 23.036us. Given that, let's see how many lines would be exposed at 1/1000 (1,000us)...

The line at the bottom of the exposed stripe has just opened immediately before the flash, and has the higher exposure. The line above it was opened 23us earlier and so on. At some point, we find a line that was opened 1.000us earlier. That line's shutter closed just before the flash, so it's not exposed. (Anyway, the result would be 1,000/23.036 or 43.4 lines.)

So, the probability is that 43.4 lines would be exposed on any given flash, though we don't know which ones. (Fewer lines are exposed if the flash happens during the "dead time" between the last and first lines.)

Mark's work also "illuminates" the amount of rolling shutter for this camera. The sensor reads 1080 lines, and there are essentially 367 "virtual lines" of dead time, where no line is read. (43410/30 - 1080 = 367) That gives a rolling shutter effect of 74.6%.

What does this mean? Let's say you are doing a horizontal pan, and let's say that the top pixel of a telephone poll moves by 100 pixels per frame. In this example the bottom pixel of the telephone poll will be offset by 75 pixels compared to vertical, when shooting with the 5D MkII.

I would guess that the D90 has a rolling shutter effect of near 100%. Crop the 5D MkII to 720p, and it would have a rolling shutter of 50% (75% x 720/1080 = 50%) So, don't scale your sports videos for Vimeo - crop them!

Anyway, congratulations on your test, Mark. I think it's conclusive. The only error would be if your flash is a bit slow. Judging from the hard edges on the pics, that's not the case. It's fast enough for this test. (Soft edges would indicate a non-instantaneous flash duration.)

On my test, I used the shutter times of the Canon in photo mode as a reference. My guess is that the camera's still mode shutter times are a bit faster than published. I would guess that photographers are more likely to complain if the shutter is too slow than too fast, and Canon built in some margin.

BTW Mark, if you get a chance, repeat your test at 1/30. I'd like to know if I should avoid 1/30 because it's too slow, or if the result is similar to 1/40 and 1/50 displayed.

Congratulations on a great test and results!

Mark Hahn · March 30th, 2009, 10:04 PM

Quote:

Originally Posted by Bernard Racelis

If the amount of time it takes to sweep the whole frame is proportional to the shutter/exposure which I think you were suggesting,
then if the shutter were 1/2000 and the whole frame can be read as fast as 1/2000 or so, then we wouldn't see as much skew / wobble.

Aha! I think I understand our misunderstanding now. I definitely don't think the sweep time of the frame varies. Quite the opposite. The rolling speed is a fixed lines/second and there are a fixed number of lines in a frame so the sweep time is constant.

The way I like to think of it is that there is a row reset that continually moves down the screen. Each row is totally emptied before the next one starts emptying. It starts at the top every 1/30 of a second and moves down at the rolling rate (how fast rows go by).

If exposure is 1/100, then 1/100 second after the reset row starts at the top, the reading row starts at the top. It then moves down the screen, completely reading each row before starting on the next one. It moves at the exact same rolling rate as the erase rolling rate.

Thus, every single pixel in the frame has exactly 1/100 second of time between the erase going by and the reading going by. It is being exposed during that time, which is 100th of a second.

Since the exposure time is the difference in time of the start of each type of row operation, that will directly affect the height of the band. If they were to start at the same time, the band would be zero high and the exposure time would be zero. If the reading line started 10 rolling rate units of time later than the erase started, the band would be 10 lines high and the exposure time would be ten times the (inverse of) the rolling rate.

So I imagine the reset doing the same thing no matter what, and the reading line happening at the right time for the exposure. Am I making any sense?

Mark Hahn · March 30th, 2009, 10:31 PM

Quote:

Originally Posted by Jon Fairhurst

I think the analysis is correct (and brilliant, BTW.)

Thanks mucho, but I was just expanding on your work.

Quote:

Assume that the flash of light is infinitely brief. At a given point of time, some sensors are held in reset (not being exposed), and others are being exposed at ambient (darker) levels.

This is just a guess of course, but I'd wager that all pixels are always being exposed. It doesn't matter what is behind the reading line so it just lets them keep being exposed.

Quote:

On my test, I used the shutter times of the Canon in photo mode as a reference. My guess is that the camera's still mode shutter times are a bit faster than published.

That could be true. I didn't see your photos, but I'd imagine the blur I showed at the beginning of the first post would be hard to measure since the ramp up at the beginning and then ramp down at the end.

Quote:

BTW Mark, if you get a chance, repeat your test at 1/30. I'd like to know if I should avoid 1/30 because it's too slow, or if the result is similar to 1/40 and 1/50 displayed.

I think 1/30 is impossible because that is all the lines including the ones "off the screen". I would always just get some 70% of the band. So I'm guessing 1/50 is as slow as it gets. I didn't show the picture but its band is most of the screen.

Oh, I realize you want me to verify what I just said by actually seeing 1/50 when it is reading 1/30. I'll do that.

I think I'll repeat the whole range. It is very easy to shoot, very easy to measure the frames, and I'll make sure I see the same results. I'll go down to 1/30 and up to as high as I can. You'll note that the error gets higher as the speed is faster and the number of lines decrease.

P.S. When I first brought the video up in premiere, I couldn't find the exposed frames. They were needles in a haystack. I was really stuck until I noticed the audio track had little tick marks on it. Whenever the flash went off, it made a popping sound which showed up as the tick. I just went tick to tick to see the frames. It was fun to notice that the exposed frame was a tiny bit before the sound tick. It took a while for the sound to get to the microphone.

Jon Fairhurst · March 30th, 2009, 11:41 PM

Quote:

Originally Posted by Mark Hahn

It was fun to notice that the exposed frame was a tiny bit before the sound tick. It took a while for the sound to get to the microphone.

Hey, this is another interesting test. We know that sound travels at about 1ms per foot. (Light is just a wee bit faster...) If you measure the time between the exposure and the sound - as well as the distance between the flash and the mic, you'll be able to measure the accuracy of the sound sync in the camera.

Mark Hahn · March 30th, 2009, 11:51 PM

Quote:

Originally Posted by Jon Fairhurst

Hey, this is another interesting test. We know that sound travels at about 1ms per foot. (Light is just a wee bit faster...) If you measure the distance between the exposure and the sound - as well as the distance between the flash and the mic, you'll be able to measure the accuracy of the sound sync in the camera.

Why am I the one who has to do all this work? :-)

I have taken a new video series and just started measuring the results.

After I saw your question above, I expanded the timeline to where the sound tick could be compared to the frames. This time I was holding the flash next to the mic and I saw no difference between sound and frames. When the flash would show up in two frames the tick was actually between the frames. So it appears to be dead-on, at least within the 1/30 sec of error margin. This video was opened directly in premiere so only the latest quicktime codec was involved.

Mark Hahn · March 31st, 2009, 12:28 AM

I've only processed the 1/30 and to my surprise it had a slower speed than 1/50. The band covered 1309 lines (out of 1447) which I had to figure out from adjacent frames since that is more than 1080. I saw many frames which were fully exposed from the flash.

1309 lines would be an exposure rate of 1/33. This makes quite a bit of sense if you think they designed it to get as close to 1/30 as they possibly could. They only missed by 10%.

If it can do 1/33, why did it do 1/50 when the read-out was 1/40. Weird.

I'm going to redo 1/30, 1/40, and 1/50 in multiple recordings to see if I can get them to vary.

March 30th, 2009, 03:48 PM	#25
Jon Fairhurst Inner Circle Join Date: May 2006 Location: Camas, WA, USA Posts: 5,513	Mark, Bernard, I think the analysis is correct (and brilliant, BTW.) It's kind of the quantum physics of sensor analysis - it uses probability, in a way. Assume that the flash of light is infinitely brief. At a given point of time, some sensors are held in reset (not being exposed), and others are being exposed at ambient (darker) levels. The exposure rolls from line to line. When the exposure is long, many lines are being exposed at the same time. When the exposure is short, few lines are being exposed simultaneously. Then BAM, the flash happens, and we get to see exactly how many lines were open simultaneously. Mark shows the delay between lines to be 23.036us. Given that, let's see how many lines would be exposed at 1/1000 (1,000us)... The line at the bottom of the exposed stripe has just opened immediately before the flash, and has the higher exposure. The line above it was opened 23us earlier and so on. At some point, we find a line that was opened 1.000us earlier. That line's shutter closed just before the flash, so it's not exposed. (Anyway, the result would be 1,000/23.036 or 43.4 lines.) So, the probability is that 43.4 lines would be exposed on any given flash, though we don't know which ones. (Fewer lines are exposed if the flash happens during the "dead time" between the last and first lines.) Mark's work also "illuminates" the amount of rolling shutter for this camera. The sensor reads 1080 lines, and there are essentially 367 "virtual lines" of dead time, where no line is read. (43410/30 - 1080 = 367) That gives a rolling shutter effect of 74.6%. What does this mean? Let's say you are doing a horizontal pan, and let's say that the top pixel of a telephone poll moves by 100 pixels per frame. In this example the bottom pixel of the telephone poll will be offset by 75 pixels compared to vertical, when shooting with the 5D MkII. I would guess that the D90 has a rolling shutter effect of near 100%. Crop the 5D MkII to 720p, and it would have a rolling shutter of 50% (75% x 720/1080 = 50%) So, don't scale your sports videos for Vimeo - crop them! Anyway, congratulations on your test, Mark. I think it's conclusive. The only error would be if your flash is a bit slow. Judging from the hard edges on the pics, that's not the case. It's fast enough for this test. (Soft edges would indicate a non-instantaneous flash duration.) On my test, I used the shutter times of the Canon in photo mode as a reference. My guess is that the camera's still mode shutter times are a bit faster than published. I would guess that photographers are more likely to complain if the shutter is too slow than too fast, and Canon built in some margin. BTW Mark, if you get a chance, repeat your test at 1/30. I'd like to know if I should avoid 1/30 because it's too slow, or if the result is similar to 1/40 and 1/50 displayed. Congratulations on a great test and results! __________________ Jon Fairhurst

March 28th, 2009, 04:09 AM	#16
Mark Hahn Major Player Join Date: Dec 2008 Location: Laguna Niguel, CA Posts: 277	(I apologize in advance for all these posts. I shouldn't be using the forum to "think out loud"). I found a good explanation (similar to my last theory) here ... Point Grey Research (PGR) - Support - Knowledge Base "The time delay between a row being reset and a row being read is the integration time. By varying the amount of time between when the reset sweeps past a row and when the readout of the row takes place, the integration time can be controlled. Since the integration process moves through the image over some length of time, some motion blur may become apparent". Then he used a fan for an example (what a coincidence) ... Note that the blade of the fan is stretched longer, not shorter. This makes the values in my original post start to make some sense. The faster number in each result is probably correct. I need to figure out how to use my current images, or to change my setup, to correctly determine the real shutter speed to some accuracy, which was my original goal. I think I understand things well enough now to try that.

March 28th, 2009, 03:05 PM	#19
Mark Hahn Major Player Join Date: Dec 2008 Location: Laguna Niguel, CA Posts: 277	OK, I've got some results from the flash experiment. That sure was easy. The setup was just the camera pointing at a wall and me flashing the strobe manually. I could see the band clearly on the LCD on every flash so I kept flashing until I caught one where the band was wholly contained on the screen. This turned out to be quite hard for slow speeds. At 1/40 it was usually off the top or bottom or wrapped around with some on the top and some on the bottom. I had to take dozens of flashes to get one in the screen. Here are a few pics at 1/60, f2, ISO 100. The left band, which is only partially visible, is 362 scan lines (33% of the height) and the right band is 435 scan lines (40%). Because the left is shorter, we know there is idle time of at least 7% where there is no scan happening. The height of the full band was less than 50%. The dead time would explain this because the rolling would have to be faster than 1/30 sec from top to bottom. If the exposure really was 1/60, then that would make the rolling rate 26,100 rows/sec (435*60). This is 38.3 usecs per row. The flash time of my 430ex is rated at 1.2ms or less, which might be a problem. I set it for lowest power which I'm hoping would be the shortest time. We'll see if the numbers indicate a problem. If this is all correct I can now determine the exact exposure time using this rolling rate and the number of lines exposed. Let's check this against some other speeds. 1/40 869 1/30.03 1/50 885 1/29.49 1/60 435 1/60 (duh) 1/80 436 1/59.8 1/125 338 1/77.2 (?) The flash speed doesn't seem to be a problem because the 1/50 and 1/60 numbers were almost exactly 2 to 1. But once again I have a problem. I am getting 1/30 readings which is impossible because I know there is dead time. This tells me 1/60 isn't really 1/60. This would also explain the weird 1/77. So what is the real rolling rate? How can I figure it out if I can't trust any of the readings? I need to do some more thinking. Any ideas? P.S. Jon and my results are in agreement that 1/40 and 1/50 are the same.

March 28th, 2009, 04:23 PM	#20
Mark Hahn Major Player Join Date: Dec 2008 Location: Laguna Niguel, CA Posts: 277	I've cracked the code. On the 1/40 video (speed doesn't matter) I knew there were 869 lines in an exposure. I found a pair of adjacent frames where the bottom was exposed on the first and the top on the second. I added up the heights of each band and subtracted that from 869 and found the number of "lines" in the idle time. Adding this number to 1080 and multiplying by 30 gave me a rolling rate of 43,410 rows/sec. I recalcuated the results I measured before using this rate. The numbers on the left are what I suspect are the actual values and the right side are the exact measurements. The difference from the actual numbers is negligable. 1/40 => 1/50 (869 => 1/49.95) 1/50 => 1/50 (885 => 1/49.05) 1/60 => 1/100 (435 => 1/99.79) 1/80 => 1/100 (436 => 1/99.56) 1/125 => 1/125 (338 => 1/128.4) So It is simple. In this range it is either 1/50, 1/100, or 1/125. The 125 is a bit weird but I think it is real. I might measure some higher ones but I don't care. If anyone cares feel free to beg for me to shoot some more tests. I think my work is done here. I believe these numbers are correct because they are so clean and logical. They are usable when designing any scheme for controlling that camera. I now know I can use either 1/40 or 1/50 and get 1/50, which not only gives repeatable results between scenes, but is very close to my holy grail of 1/48. Others who want faster speeds can use either 1/60 or 1/80 and always get 1/100. Jon: I'm sorry but your holy grail of 1/60 is impossible.

March 30th, 2009, 03:29 PM	#24
Bernard Racelis Regular Crew Join Date: Dec 2007 Location: New York City Posts: 82	Mark, It could also be that I did not understand exactly the method that you were using to perform the test -- I hadn't bothered reading all the posts in detail since you've said that this thread has become cluttered. Perhaps you can explain in a simple/short way what exactly you're measuring, without all the clutter.

March 31st, 2009, 12:28 AM	#30
Mark Hahn Major Player Join Date: Dec 2008 Location: Laguna Niguel, CA Posts: 277	I've only processed the 1/30 and to my surprise it had a slower speed than 1/50. The band covered 1309 lines (out of 1447) which I had to figure out from adjacent frames since that is more than 1080. I saw many frames which were fully exposed from the flash. 1309 lines would be an exposure rate of 1/33. This makes quite a bit of sense if you think they designed it to get as close to 1/30 as they possibly could. They only missed by 10%. If it can do 1/33, why did it do 1/50 when the read-out was 1/40. Weird. I'm going to redo 1/30, 1/40, and 1/50 in multiple recordings to see if I can get them to vary.