I kinda hate this article because it could be so much better and yet it isn’t.
The headline is clearly, in some sense, silly. It’s silly because the usual notion people have of fps is something like video where you need to move your roll of film or read out from your sensor at that rate. The ‘frame rate’ for a photo finish is much more incidental to the way the image is made.
If you have an image from a typical digital camera that is 4000px high formed by a typical two-curtain shutter that shutters in 10ms, you expose 400 rows per ms or, in some silly sense, 400k FPS. If you wanted a 1-pixel high slit to be exposed then that’s an exposure of 1/400000s, which is faster than any camera I’m aware of. But I think the analogy is useful – instead of a shutter moving a slit over the focal plane, exposing different parts of the sensor at slightly different times, imagine the sensor being moved behind a fixed slit and then reading out the values from it. You can reasonably easily imagine doing this with film too – just move it past the slit at a steady rate – not dissimilarly from taking an old-school panorama camera and moving the body instead of the lens.
I think the thing happening here is not a moving sensor but rather a 1 pixel wide sensor (or perhaps a few pixels for colour reasons). This makes it thousands of times smaller than the resolution of the final image so even a fairly typical cmos sensor at 2e9 pixels per second could read 50000 pixel ‘frames’ at ‘40k FPS’. (In practice the number would probably be lower for synchronisation reasons). When your frame is very skinny, that still gives you plenty of resolution.
I don’t like the article because they did some silly arithmetic that produces a big number instead of digging into interesting details, e.g.
- talking about how the system works as a whole (when does it decide to start/stop the image, maybe something about buffering)
- talking about how much light you need and how you get enough
- talking about the optics, how you keep everyone sharp while still getting enough light, how you even focus such a thing
- talking how you make sure the camera is setup fairly (eg perpendicular to the lanes, able to get a good view of all the lanes)
- maybe something about reliability and how you avoid the bad scenario of the system failing when it matters most
> You can reasonably easily imagine doing this with film too – just move it past the slit at a steady rate – not dissimilarly from taking an old-school panorama camera and moving the body instead of the lens.
No need to imagine, this is how photo finish cameras at horse and greyhound races have worked for decades.
I agree so much. Not a single mention of what a strip camera is and how the picture we see is not a picture but a composite of 1px wide timestamped slices.
Shame on Petapixel
Photo finishes are typically done with a line scan camera. It only captures a single column of pixels at a time. So the horizontal axis in the image is actually time, not space. Super cool stuff.
Outside sports, digital line scan cameras are used in various quality control applications (objects on conveyer belts, vehicle mounted road/rail scanners, etc). This unit can film an 8k px strip at a rate of 80khz.
There is a similar technique in astronomy, though much slower, called drift-scan imaging. Typically when you take a long exposure astronomical image the telescope has to rotate to track the star as it moves across the sky. So in traditional imaging you track the object, make your exposure, and then when the exposure is done you read out the image. The downside is that while you're reading out the image you can't do anything else and for an astronomical CCD it can take on the order of a minute or so. So you lose ~5% of your observing time just to reading out images. (It's more if you have to slew to different locations on the sky.)
In drift-scan imaging you keep the telescope pointed at a fixed location and you continuously read out the image at a rate that matches the motion of the stars across the field of view. This allows you to continuously collect imaging from a strip across the sky.
Oh yeah I took that picture on the Wikipedia article! I have a couple more on my website [1]. One of these days I want to go to Atherton station with my line scan camera to scan some Caltrains.
Fun fact! This is actually how many earth observation satellites work too! Except the motion comes not from the subject, but the satellite orbit itself. It's called a pushbroom camera.
To the question of the Omega logo and Olympic rings asked below, this article states "In fact, the entire photo is literally the finish line viewed one pixel at a time (the Omega and Olympic branding is added at the top)."
I had to combine Tuna-Fish comments and yours to understand what's going on.
> And the Omega/Olympics flag "banner" behind the runners is not a banner, it is a led screen, one pixel wide.
So the one pixel wide scan line camera outputs and image that has time as horizontal axis, and the one pixel screen that draws the omega banner is a 40kfps single line video that does marketing one the photo finish that likely serves as a clock sync check.
Ah I was wondering why you can't actually see the finish line in the image. This wasn't really described well. Are the shadows of the runners artificial then?
Because this is a line camera, the entire image is the finish line. That is, each vertical column of pixels is what was on the finish line at that particular point in time.
That’s also the reason for the distortions. It’s not a single frame taken at one time.
I was there at the time, and I could clearly see how the advertising board works. It's a vertical line of LEDs that constantly rotates through the columns of the logo. To the naked eye it just seems to be flickering randomly.
According to Reddit thread on this, the led board behind is specifically animated with the right timing so the content shows up correctly in the line camera.
When I read "line scan camera" I thought of "rolling shutter" which in digital cameras works out to be a horizontal line at a time, which causes interesting artifacts with quick moving objects like propellers or windshield wipers.
Questions that came to mind and answers as far as I can tell:
Q: Are the scan lines parallel to the plane of completion?
A: There is only one scan line and it is parallel to the plane of completion. "In track, the cameras are only focused on the 5 mm near the finish line." [0]
Q: If yes, is the sequence of lines scanned in the same direction as the competitor movement?
A: There is only one scan line.
Note that the rolling shutter effect is not limited to digital photography: it also affects physical focal plane shutters at high shutter speeds due to the time it takes for shutter curtains to move across the frame.
This effect is the origin of the "fast race car slants forward" trope:
First I was wondering why there are ad banners visible in the image. With a line scan camera the background which is standing still should actually only have horizontal stripes.
But since the company Omega produced this photo it actually makes sense to artificially add their logo next to the Olympic rings.
i read that that actually a led screen that basically is synced to the camera so it appears as a 2d banner when in actuallity its a 1d banner that changes over time which when displayed like it is looks like a 2d banner. kinda neat. In a weird way similar to the the adverts painted on the pitch or trackside that look like 2d on screen dogs but only look that way from a certain camera angle
It’s slightly misleading to say it’s time rather than space, though, because of course the subjects are moving and so each column does correspond to a different position (with respect to each one). With respect to the track, however, moving horizontally is obviously just moving in time.
Another weird (until you think about it) thing that’s true: every pixel in the image is at the finish line.
The 'vertical' axis is spatial, representing position across the track, but I would argue that if we take your final sentence and append "...at sequential time intervals" to it, we end up making the case that the horizontal axis in the image is actually time, not space. Each column is the finish line at a different time, and perhaps counter-intuitively (given the usual convention), the right-most column is the earliest.
This is clearer, I think, from considering how the banner is made: as others have pointed out, it is produce by having a single row of LEDs in line with the finish line, each one being turned on and off in a sequence which results in the banner shown in the picture.
Stray gusts of wind, slight irregularities in the track's surface, other tiny inequalities between lanes or starting positions are probably going to nudge runner performance enough to overwhelm the precision of even a fairly unremarkable finish camera.
If fairness were the primary consideration, then they would use any old junk camera and be generous in calling ties. This kind of theatrical accuracy is there purely so that Omega can advertise.
One could say that 1st place only won because some random factor favored him. But one could just as easily say that 2nd place was only so close because some random factor favored him. If there is no way to measure or control a factor, there is no way to know whom it favored.
So race organizers try to make things as fair as possible, but everything beyond that is just “part of the game.” The athletes accept that, and spectators should as well.
I agree with your factual claims, many things can influence the result.
But I don't quite see how using less accurate measuring tools makes sense. At the end of the day getting across the finish line first is the only thing that really matters. Seems to me that more accuracy can only be a good thing
Do they go back to ensure everyones foot is on the block to a precise mm, or to ensure that the runner didn't flinch or move their foot from the block 1/1000 of a second early before the sound hit their ears, taking into account the distance from each runner's ear to the closest speaker playing the starting sound?
The extreme level of precision, I suspect, is not the same at the finish and the start.
The sprinters have speakers near their starting blocks that broadcast the start noise. If this seems excessive, consider that a starting pistol noise takes longer to reach the outer lanes from the inner ones than was the margin between first and second place in this race.
This just illustrates that this specific competition has become meaningless. It might measure their auditory reaction time more than their sprinting speed.
If a races are decided by such small amounts, like lane 1 being a millimeter ahead than lane 2, then the lanes have to be the same down to a millimeter for it to be a fair race. I wonder if timing will outpace feasible construction before long and we’ll have to say, “lane 1 finished a zillionth of a second ahead, but that makes it a tie because we can’t tell if they started a zillionth of a meter closer.”
That's already the case in swimming, as I commented a few days ago.
It's not possible to build swimming pools to tolerances such that timing to 1/1,000th of a second would be fair (that is, based on athletic performance rather than distance traveled), so swimming events are only judged to the nearest 1/100th of a second. Finishers within the same 1/100th are considered to have tied.
I know it’s vanishingly rare, but in the hypothetical where one swimmer finishes at X.004 and another finishes at X.005, does that mean they will get scored as X.00 and X.01 and not tie despite being only 1/1000th of a second different? I know for normal meets that’s what would happen, but didn’t find an answer in the FINA standards doc for the olympics/world championships.
As someone who worked closely to the Omega / Swisstiming operations at the Olympics, this is super cool stuff. Congratulations to all involved in being able to deliver when it mattered.
The team also makes really interesting stuff on other sports, such as Beach Volley. Worth checking how it's done.
There was controversy in 2012 at the U.S. Olympic Trials when Allyson Felix & Jeneba Tarmoh finished neck and neck -- basically the cameras at the time couldn't make out the difference [0]. I think that was pretty hard on the athletes, so I'm glad to see we've made some progress.
Back in 1987 I started my first job at the Australian DoD. Our group's remit was explosives testing and instrumentation. One of my jobs was to program a hardware controller/sequencer[0] that would trigger numerous equipment: detonators, accelerometer recorders and high-speed analogue cameras. At our facility, we had two bunkers (for in-situ explosives testing), that had a very small window (maybe 15x15cm that was about a foot deep), behind which sat a very big, very expensive, and very fast high-speed camera. That thing could do 10^6 FPS for very short bursts. It had a relatively long lead time to spin up to speed (hence the need for sequencers to precisely kick of the camera and then detonate the subject device at the right moment). I vaguely remember being told that it cost several million dollars (pretty expensive back in 1987). The team that were in charge of that thing also had a couple of Silicon Graphics IRIS workstations, while I was stuck button punching codes on a vertical, rack-mounted keyboard for the sequencer. That camera and the graphics workstations felt like we were living in the future.
I've tried to find out more about the camera but honestly all I can recall is how fast it was, how expensive it was, and that the whole rig took about a square meter or two of floor space. But to eventually be able to see an explosion evolving frame by frame was pretty cool. I fully expect given the pace of change, that by the time 40 years has passed, what was once DoD specced "secret" tech, will be consumer level tech - shrunk from m^3 to several mm^3, and cost down by a factor of 10^5. Amazing.
Readit News
The headline is clearly, in some sense, silly. It’s silly because the usual notion people have of fps is something like video where you need to move your roll of film or read out from your sensor at that rate. The ‘frame rate’ for a photo finish is much more incidental to the way the image is made.
If you have an image from a typical digital camera that is 4000px high formed by a typical two-curtain shutter that shutters in 10ms, you expose 400 rows per ms or, in some silly sense, 400k FPS. If you wanted a 1-pixel high slit to be exposed then that’s an exposure of 1/400000s, which is faster than any camera I’m aware of. But I think the analogy is useful – instead of a shutter moving a slit over the focal plane, exposing different parts of the sensor at slightly different times, imagine the sensor being moved behind a fixed slit and then reading out the values from it. You can reasonably easily imagine doing this with film too – just move it past the slit at a steady rate – not dissimilarly from taking an old-school panorama camera and moving the body instead of the lens.
I think the thing happening here is not a moving sensor but rather a 1 pixel wide sensor (or perhaps a few pixels for colour reasons). This makes it thousands of times smaller than the resolution of the final image so even a fairly typical cmos sensor at 2e9 pixels per second could read 50000 pixel ‘frames’ at ‘40k FPS’. (In practice the number would probably be lower for synchronisation reasons). When your frame is very skinny, that still gives you plenty of resolution.
I don’t like the article because they did some silly arithmetic that produces a big number instead of digging into interesting details, e.g.
- talking about how the system works as a whole (when does it decide to start/stop the image, maybe something about buffering)
- talking about how much light you need and how you get enough
- talking about the optics, how you keep everyone sharp while still getting enough light, how you even focus such a thing
- talking how you make sure the camera is setup fairly (eg perpendicular to the lanes, able to get a good view of all the lanes)
- maybe something about reliability and how you avoid the bad scenario of the system failing when it matters most
No need to imagine, this is how photo finish cameras at horse and greyhound races have worked for decades.
Which is a fairly big deal as sometimes it isn't set-up correctly.
Who won the Amstel Gold Race? Human error in photo-finishes (April 29, 2021 ) [1]
[1]: https://news.ycombinator.com/item?id=26970854
Deleted Comment
https://en.wikipedia.org/wiki/Strip_photography
Early film versions used a highspeed spinning slit aperture to film fast objects. This paper from 1931 shows some very impressive results for ballistics, including the shockwave: https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.1931...
Outside sports, digital line scan cameras are used in various quality control applications (objects on conveyer belts, vehicle mounted road/rail scanners, etc). This unit can film an 8k px strip at a rate of 80khz.
https://www.youtube.com/watch?v=OXUwJOJ7fMk
In drift-scan imaging you keep the telescope pointed at a fixed location and you continuously read out the image at a rate that matches the motion of the stars across the field of view. This allows you to continuously collect imaging from a strip across the sky.
[1] https://daniel.lawrence.lu/photos/
For a while, handheld scanners were a thing: hand-held strip cameras with roller wheels you would swipe across a page.
To the question of the Omega logo and Olympic rings asked below, this article states "In fact, the entire photo is literally the finish line viewed one pixel at a time (the Omega and Olympic branding is added at the top)."
> And the Omega/Olympics flag "banner" behind the runners is not a banner, it is a led screen, one pixel wide.
So the one pixel wide scan line camera outputs and image that has time as horizontal axis, and the one pixel screen that draws the omega banner is a 40kfps single line video that does marketing one the photo finish that likely serves as a clock sync check.
That’s also the reason for the distortions. It’s not a single frame taken at one time.
As the runners' shadows cross the finish line, they are recorded by the camera in the same way that the runners are.
The shadows are what's projected on the line from the stadium lights, and the shape is how they changed in time.
This short video might be a better explanation of the "line camera" concept of multiple photos of a single line being stretched out over the x-axis :
https://www.youtube.com/watch?v=Ut0nKdLCAEo&t=0m23s
You can see it here at 9m50s:
https://youtu.be/7Xnr805bm4E?feature=shared&t=590
It's just a single animated strip, one pixel wide. I assume the camera array is on the opposite side.
Deleted Comment
Deleted Comment
Questions that came to mind and answers as far as I can tell:
0. https://www.axios.com/2024/08/05/noah-lyles-wins-gold-track-...Also interesting: https://www.hodinkee.com/articles/olympics-2024-mens-100m-ph...
This effect is the origin of the "fast race car slants forward" trope:
https://about.usps.com/news/national-releases/2011/images/pr...
https://imsmuseum.org/adoptable_car/1911-winner/
But since the company Omega produced this photo it actually makes sense to artificially add their logo next to the Olympic rings.
Another weird (until you think about it) thing that’s true: every pixel in the image is at the finish line.
This is clearer, I think, from considering how the banner is made: as others have pointed out, it is produce by having a single row of LEDs in line with the finish line, each one being turned on and off in a sequence which results in the banner shown in the picture.
If fairness were the primary consideration, then they would use any old junk camera and be generous in calling ties. This kind of theatrical accuracy is there purely so that Omega can advertise.
So race organizers try to make things as fair as possible, but everything beyond that is just “part of the game.” The athletes accept that, and spectators should as well.
We don’t need to run the race.
But I don't quite see how using less accurate measuring tools makes sense. At the end of the day getting across the finish line first is the only thing that really matters. Seems to me that more accuracy can only be a good thing
The extreme level of precision, I suspect, is not the same at the finish and the start.
Deleted Comment
Deleted Comment
Deleted Comment
It is not. They do a great job of getting people to think that it is, however.
It's not possible to build swimming pools to tolerances such that timing to 1/1,000th of a second would be fair (that is, based on athletic performance rather than distance traveled), so swimming events are only judged to the nearest 1/100th of a second. Finishers within the same 1/100th are considered to have tied.
<https://news.ycombinator.com/item?id=41128437>
<https://olympstats.com/2014/02/12/timing-accuracy-at-the-oly...>
FINA standards: <https://resources.fina.org/fina/document/2022/02/08/77c3058d...> (PDF).
The team also makes really interesting stuff on other sports, such as Beach Volley. Worth checking how it's done.
Dead Comment
[0] https://www.theguardian.com/sport/2012/jul/02/jeneba-tarmoh-...
[0] Zilog Z8 BASIC model controller.