Don't keep your head in the game
October 30, 2020 3:11 AM Subscribe
This is gold.
posted by Literaryhero at 4:35 AM on October 30, 2020
posted by Literaryhero at 4:35 AM on October 30, 2020
My guess is that next week all the captchas will be "Click all the frames than include a ball".
posted by Harald74 at 5:25 AM on October 30, 2020 [46 favorites]
posted by Harald74 at 5:25 AM on October 30, 2020 [46 favorites]
And y'all want self-driving cars.
posted by GamblingBlues at 5:36 AM on October 30, 2020 [43 favorites]
posted by GamblingBlues at 5:36 AM on October 30, 2020 [43 favorites]
Stupid camera failure isn't always funny. This one, for example, cost an estimated £200 million.
posted by Cardinal Fang at 5:56 AM on October 30, 2020 [2 favorites]
posted by Cardinal Fang at 5:56 AM on October 30, 2020 [2 favorites]
Is the AI camera operator mechanically operating a camera, or is it taking the feed from a wide shot and digitally panning and zooming? As I read the story I assumed the former, but watching the video it really looks like that later.
posted by thecjm at 6:24 AM on October 30, 2020
posted by thecjm at 6:24 AM on October 30, 2020
This title isn't misleading but really understates the baldness and whiteness of the man's head. A lot appears to be due to the glare from the sun, but this isn't necessarily bad computer vision at play. If you take a screenshot of the game and compare the ball to the man's head it is eerily similar. To the point of asking yourself, how do you as a person tell the difference?
This particular problem could have been solved if they simply used person recognition to filter out the people. This would have detected the person and then segmented the head as a feature and created a mask on the refer, eliminating it entirely. That's computationally expensive. There's a way to do it using KLT tracking to reduce the load but you must remember with anything that's related to computer vision your'e doing this how many FPS, so your baseline is around 33.3 ms if your'e at 30 fps. That means anything you're doing has to take place within 333 ms.
There's also another occluding algorithms out there. If the ball is on a set of a coordinates and is partially or entirely out of view of the camera it assume it is on the same trajectory and each frame it is occluded the chance of it being in the predicted area is reduced. I'm guessing the ball is occluded to some degree and the algorithm decided the head was the best match of what a ball is (with a much lower confidence) instead of assuming the ball didn't travel halfway across the field and was simply occluded. This is a real problem when talking about people tracking in a store or public space where someone walks behind someone. Much worse than a ball on a soccer field you have 50 balls (people) in frame all weaving in and out how do you track them?
Really all sport vision technology tackles some of the easiest problems in the computer vision domain. You have fields with clearly marked boundaries that make calibration incredibly easy compared to any other 3D space or volume. The amount of people in play are known everything that needs to be tracked is brightly colored or separated by uniform. A lot of sports even have numbering on the uniforms. Computer vision isn't magic, it tracks based on the same rules you or I would unconsciously track something. The hard part is picking which rules work in the small amount of time you have to process it.
In the future, all the referee needs to do is wear a hat.
posted by geoff. at 6:42 AM on October 30, 2020 [12 favorites]
This particular problem could have been solved if they simply used person recognition to filter out the people. This would have detected the person and then segmented the head as a feature and created a mask on the refer, eliminating it entirely. That's computationally expensive. There's a way to do it using KLT tracking to reduce the load but you must remember with anything that's related to computer vision your'e doing this how many FPS, so your baseline is around 33.3 ms if your'e at 30 fps. That means anything you're doing has to take place within 333 ms.
There's also another occluding algorithms out there. If the ball is on a set of a coordinates and is partially or entirely out of view of the camera it assume it is on the same trajectory and each frame it is occluded the chance of it being in the predicted area is reduced. I'm guessing the ball is occluded to some degree and the algorithm decided the head was the best match of what a ball is (with a much lower confidence) instead of assuming the ball didn't travel halfway across the field and was simply occluded. This is a real problem when talking about people tracking in a store or public space where someone walks behind someone. Much worse than a ball on a soccer field you have 50 balls (people) in frame all weaving in and out how do you track them?
Really all sport vision technology tackles some of the easiest problems in the computer vision domain. You have fields with clearly marked boundaries that make calibration incredibly easy compared to any other 3D space or volume. The amount of people in play are known everything that needs to be tracked is brightly colored or separated by uniform. A lot of sports even have numbering on the uniforms. Computer vision isn't magic, it tracks based on the same rules you or I would unconsciously track something. The hard part is picking which rules work in the small amount of time you have to process it.
In the future, all the referee needs to do is wear a hat.
posted by geoff. at 6:42 AM on October 30, 2020 [12 favorites]
When fans come back, the bald ones are all going to draw ball patterns on their heads.
posted by cardboard at 6:43 AM on October 30, 2020 [3 favorites]
posted by cardboard at 6:43 AM on October 30, 2020 [3 favorites]
Silly geoff., hats are for goalies.
posted by Exceptional_Hubris at 6:44 AM on October 30, 2020 [1 favorite]
posted by Exceptional_Hubris at 6:44 AM on October 30, 2020 [1 favorite]
This title isn't misleading but really understates the baldness and whiteness of the man's head.
It's like how much more bald and white could it be? And the answer is none. None more bald and white.
posted by any portmanteau in a storm at 7:15 AM on October 30, 2020 [14 favorites]
It's like how much more bald and white could it be? And the answer is none. None more bald and white.
posted by any portmanteau in a storm at 7:15 AM on October 30, 2020 [14 favorites]
Can't they just put a chip in the ball?
posted by Melismata at 7:33 AM on October 30, 2020 [2 favorites]
posted by Melismata at 7:33 AM on October 30, 2020 [2 favorites]
In the future, all the referee needs to do is wear a hat.
This isn't a problem with the referee
posted by niicholas at 7:35 AM on October 30, 2020 [10 favorites]
This isn't a problem with the referee
posted by niicholas at 7:35 AM on October 30, 2020 [10 favorites]
So often, referees aren't given the recognition they deserve.
posted by Capt. Renault at 7:37 AM on October 30, 2020 [8 favorites]
posted by Capt. Renault at 7:37 AM on October 30, 2020 [8 favorites]
Now instead of a bald man's head imagine a wedding party. And instead of a camera, imagine a Predator drone carrying Hellfire missiles.
posted by Nelson at 7:43 AM on October 30, 2020 [15 favorites]
posted by Nelson at 7:43 AM on October 30, 2020 [15 favorites]
What bawheid thought this was a good cost saving strategy? I mean, you could practically have any heid-the-baw Caley Thistle fan (either one of them) pay you to run the camera. Football is so ingrained in Scottish culture that "point the camera at the ball" is a reflex of every creature evolved past newts (sorry, Rangers fans, you're out).
posted by scruss at 7:48 AM on October 30, 2020 [8 favorites]
posted by scruss at 7:48 AM on October 30, 2020 [8 favorites]
Guys, I don't want to make a big deal of this, but I think that geoff. might be a robot.
posted by Literaryhero at 7:49 AM on October 30, 2020 [12 favorites]
posted by Literaryhero at 7:49 AM on October 30, 2020 [12 favorites]
If you take a screenshot of the game and compare the ball to the man's head it is eerily similar. To the point of asking yourself, how do you as a person tell the difference?
R/TOTALLYNOTROBOTS
posted by Pyrogenesis at 7:53 AM on October 30, 2020 [3 favorites]
R/TOTALLYNOTROBOTS
posted by Pyrogenesis at 7:53 AM on October 30, 2020 [3 favorites]
I am cackling like a loon at what a crush this camera operator has on that guy. Thank you for posting this.
posted by LobsterMitten at 8:03 AM on October 30, 2020 [4 favorites]
posted by LobsterMitten at 8:03 AM on October 30, 2020 [4 favorites]
If you take a screenshot of the game and compare the ball to the man's head it is eerily similar. To the point of asking yourself, how do you as a person tell the difference?
I usually find the presence of arms and legs attached to the shiny, vaguely spherical object to be a strong indicator.
usually...
posted by Naberius at 8:04 AM on October 30, 2020 [2 favorites]
I usually find the presence of arms and legs attached to the shiny, vaguely spherical object to be a strong indicator.
usually...
posted by Naberius at 8:04 AM on October 30, 2020 [2 favorites]
reminds me of the time I took my hat off to pose next to the summit sign at Bald Knob
posted by exogenous at 8:06 AM on October 30, 2020 [2 favorites]
posted by exogenous at 8:06 AM on October 30, 2020 [2 favorites]
In the future, all the referee needs to do is wear a hat.
This could be that man's only opportunity to get some Vitamin D.
posted by biffa at 8:15 AM on October 30, 2020 [1 favorite]
This could be that man's only opportunity to get some Vitamin D.
posted by biffa at 8:15 AM on October 30, 2020 [1 favorite]
This particular problem could have been solved if they simply used person recognition to filter out the people hired a camera operator.
posted by GCU Sweet and Full of Grace at 8:17 AM on October 30, 2020 [14 favorites]
posted by GCU Sweet and Full of Grace at 8:17 AM on October 30, 2020 [14 favorites]
If you take a screenshot of the game and compare the ball to the man's head it is eerily similar. To the point of asking yourself, how do you as a person tell the difference?
I bet he's got stories. Terrible stories.
posted by mandolin conspiracy at 8:28 AM on October 30, 2020
I bet he's got stories. Terrible stories.
posted by mandolin conspiracy at 8:28 AM on October 30, 2020
Like to see the Terminator recut with all the real AI goodness...
posted by zerobyproxy at 8:52 AM on October 30, 2020 [1 favorite]
posted by zerobyproxy at 8:52 AM on October 30, 2020 [1 favorite]
This kind of thing is exactly the explanation for why all the terminator frames in every movie opening are just firing aimlessly in all directions.
posted by Drastic at 9:04 AM on October 30, 2020 [4 favorites]
posted by Drastic at 9:04 AM on October 30, 2020 [4 favorites]
Keep your eye on the bald.
posted by aniola at 9:15 AM on October 30, 2020 [6 favorites]
posted by aniola at 9:15 AM on October 30, 2020 [6 favorites]
I get it, AI. I'm attracted to bald guys too.
posted by Gray Duck at 9:30 AM on October 30, 2020 [6 favorites]
posted by Gray Duck at 9:30 AM on October 30, 2020 [6 favorites]
I'm on team AI crush.
Also not clear how this ruins the game for fans, as the camera's obsession was much more interesting to watch than that game.
posted by chavenet at 10:29 AM on October 30, 2020 [1 favorite]
Also not clear how this ruins the game for fans, as the camera's obsession was much more interesting to watch than that game.
posted by chavenet at 10:29 AM on October 30, 2020 [1 favorite]
"…though some fans saw this as a bonus given the usual quality of performance."
Ouch.
posted by JiBB at 11:52 AM on October 30, 2020 [1 favorite]
Ouch.
posted by JiBB at 11:52 AM on October 30, 2020 [1 favorite]
This is just good clean fun!
posted by Atom Eyes at 12:27 PM on October 30, 2020 [1 favorite]
posted by Atom Eyes at 12:27 PM on October 30, 2020 [1 favorite]
> I think that geoff. might be a robot.
Really? Doesn't look like anything to me.
posted by GeckoDundee at 1:50 PM on October 30, 2020 [2 favorites]
Really? Doesn't look like anything to me.
posted by GeckoDundee at 1:50 PM on October 30, 2020 [2 favorites]
It’s a mixed bag, being a bald man these days. AI lovers are diligent but exhausting.
posted by Don.Kinsayder at 2:33 PM on October 30, 2020 [5 favorites]
posted by Don.Kinsayder at 2:33 PM on October 30, 2020 [5 favorites]
BRB writing the AI sequel...
"The Camera Who Mistook The Referee's Head For A Ball"
posted by symbioid at 4:44 PM on October 30, 2020 [5 favorites]
"The Camera Who Mistook The Referee's Head For A Ball"
posted by symbioid at 4:44 PM on October 30, 2020 [5 favorites]
Actually can we get Chuck Tingle on the task of writing the novelization?
posted by the antecedent of that pronoun at 4:58 PM on October 30, 2020 [5 favorites]
posted by the antecedent of that pronoun at 4:58 PM on October 30, 2020 [5 favorites]
This particular problem could have been solved if they simply hired a camera operator.
Sorry, this solution simply isn't scalable. It may work fine when there's only one or two soccer matches a day, but what happens when there's millions of soccer matches each second? Human moderation simply won't be able to keep up.
posted by Mr.Encyclopedia at 7:57 PM on October 30, 2020 [12 favorites]
Sorry, this solution simply isn't scalable. It may work fine when there's only one or two soccer matches a day, but what happens when there's millions of soccer matches each second? Human moderation simply won't be able to keep up.
posted by Mr.Encyclopedia at 7:57 PM on October 30, 2020 [12 favorites]
Sorry, this solution simply isn't scalable. It may work fine when there's only one or two soccer matches a day, but what happens when there's millions of soccer matches each second? Human moderation simply won't be able to keep up.
Actually a real problem that's really being solved. There's a lot more amateur matches then professional matches. Specifically determining offsides objectively or in baseball a batter's box. Utilizing a referee or umpire and then a good referee umpire is actually not scalable.
I was approached to do something for children's basketball games. Coaches and/or parents wanted to know statistics and poses used by their kids. We can say it is a large problem with society but I liken it to knowing how many steps you take when you go on a walk. Sometimes the gamification of even a game is fun to look back and review. It is certainly done with video games and other less serious competitions ... so I don't always assume we get a NYT crazy article about a 9 year old being analyzed so they can get into an Ivy League squash program but simply parents and kids curious about some analytics during a game.
But anyway amateur sports are so much harder than professional sports. Cameras are not calibrated and the courts are not set. For example a high school gym might have markings for basketball at different levels (the three point shot might be marked differently for different skills) along with markings for different sports entirely. These alone are not uniform so you can't say blue markings are for teenagers and black markings are for children. Compound this with some just bad markings, like why would someone put two black three point lines?! Theoretically you could ignore the markings altogether and use homography to calculate the distance from the basket and draw virtual lines but that's another level of complexity. Plus the lines themselves are used to create a "virtual horizon" to calibrate a shot. And of course all this with a phone camera, little budget and non-players moving on the court and you have a really hard problem set.
So this is scalable, and there's a lot of money there. Even if it is $1 a game, across every game at every YMCA, you're talking billions. That's way before the crazy amount of money you could get from a major sports apparel manufacturer with a swoosh logo who wants to know how many people are wearing their shoes on a court at any given time (true story).
In any case I was presented with this and thought about it but I didn't turn it down because it was hard or not technically feasible but because after a competitive analysis there were other companies doing it better, with much better backing and attempting to create a monopoly on school sports video analytics.
* Addendum the real hard part of all this isn't really the technology it is the expectations. So NBC/CBS/ABC/ESPN put on a great high production show. They've been doing real-time graphics during sports games for the last 50 years at least, with some sort of what we would now call AI since at least 1996. They've made a lot of improvements with near unlimited funding and tightly controlled production environments. People will say they don't expect high production values and they're generally right, what they don't get is that if their iPhone camera -- despite being a relatively good camera -- doesn't capture something it is hard or near impossible to extract information. If someone gets up during a game in the bleachers in front of you and blocks your shot or you're cheering and your kid gets out of frame you lose it. Somewhat worse are amateur videographers who think they're directing Michael Jackson during the fucking Super Bowl halftime show and insist they're getting all the shot or their camera is setup right. Yeah if you're moving your camera while your kid is moving and your frame rate sucks and I have the equivalent of a blurry photo of a Loch Ness monster I can't tell if your kid is making a shot right now, you can have Stephen A Smith watching that video and unable to tell you who made the shot. And yes this is why professional games aren't done by LeBron James' mom holding a selfie stick and why we have production teams with directors.
posted by geoff. at 12:55 PM on October 31, 2020
Actually a real problem that's really being solved. There's a lot more amateur matches then professional matches. Specifically determining offsides objectively or in baseball a batter's box. Utilizing a referee or umpire and then a good referee umpire is actually not scalable.
I was approached to do something for children's basketball games. Coaches and/or parents wanted to know statistics and poses used by their kids. We can say it is a large problem with society but I liken it to knowing how many steps you take when you go on a walk. Sometimes the gamification of even a game is fun to look back and review. It is certainly done with video games and other less serious competitions ... so I don't always assume we get a NYT crazy article about a 9 year old being analyzed so they can get into an Ivy League squash program but simply parents and kids curious about some analytics during a game.
But anyway amateur sports are so much harder than professional sports. Cameras are not calibrated and the courts are not set. For example a high school gym might have markings for basketball at different levels (the three point shot might be marked differently for different skills) along with markings for different sports entirely. These alone are not uniform so you can't say blue markings are for teenagers and black markings are for children. Compound this with some just bad markings, like why would someone put two black three point lines?! Theoretically you could ignore the markings altogether and use homography to calculate the distance from the basket and draw virtual lines but that's another level of complexity. Plus the lines themselves are used to create a "virtual horizon" to calibrate a shot. And of course all this with a phone camera, little budget and non-players moving on the court and you have a really hard problem set.
So this is scalable, and there's a lot of money there. Even if it is $1 a game, across every game at every YMCA, you're talking billions. That's way before the crazy amount of money you could get from a major sports apparel manufacturer with a swoosh logo who wants to know how many people are wearing their shoes on a court at any given time (true story).
In any case I was presented with this and thought about it but I didn't turn it down because it was hard or not technically feasible but because after a competitive analysis there were other companies doing it better, with much better backing and attempting to create a monopoly on school sports video analytics.
* Addendum the real hard part of all this isn't really the technology it is the expectations. So NBC/CBS/ABC/ESPN put on a great high production show. They've been doing real-time graphics during sports games for the last 50 years at least, with some sort of what we would now call AI since at least 1996. They've made a lot of improvements with near unlimited funding and tightly controlled production environments. People will say they don't expect high production values and they're generally right, what they don't get is that if their iPhone camera -- despite being a relatively good camera -- doesn't capture something it is hard or near impossible to extract information. If someone gets up during a game in the bleachers in front of you and blocks your shot or you're cheering and your kid gets out of frame you lose it. Somewhat worse are amateur videographers who think they're directing Michael Jackson during the fucking Super Bowl halftime show and insist they're getting all the shot or their camera is setup right. Yeah if you're moving your camera while your kid is moving and your frame rate sucks and I have the equivalent of a blurry photo of a Loch Ness monster I can't tell if your kid is making a shot right now, you can have Stephen A Smith watching that video and unable to tell you who made the shot. And yes this is why professional games aren't done by LeBron James' mom holding a selfie stick and why we have production teams with directors.
posted by geoff. at 12:55 PM on October 31, 2020
the future is weird
posted by Monochrome at 1:25 PM on October 31, 2020 [2 favorites]
posted by Monochrome at 1:25 PM on October 31, 2020 [2 favorites]
« Older Age of Discord II | A Short Yikes! Newer »
This thread has been archived and is closed to new comments
I imagine it would look a lot like this.
posted by Faint of Butt at 4:02 AM on October 30, 2020 [14 favorites]