AI that mistook a bald head for a soccer ball | Hacker News Reader

AI that mistook a bald head for a soccer ball

kottke.org

162 pointssogen79 comments

Yesterday: https://news.ycombinator.com/item?id=24955651

Could be worse. An eagle once mistook Aeschylus's bald head for a rock and dropped a tortoise on it (expecting the rock to smash the tortoise shell), killing Aeschylus. https://www.goldencharter.co.uk/news-and-info/2017/six-stran....

what a tragedy!

Interesting angle for a funeral service ad..

> To add a further strange twist to the death of Aeschylus, Roman author Pliny, suggested that Aeschylus had been spending a lot of time outdoors to avoid a prophecy that he would be killed by a falling object.

I don't get it. If you're worried about being killed by a falling object, wouldn't you want to stay under the cover of a roof?

Buildings back then probably collapsed much more frequently.

The entire hellenic world was pretty much an earthquake zone too

I suspect at least part of this story originated as a joke given his status as the "father of tragedy".

It really depends whether you live in an area where earthquakes, collapsing houses, things falling from high shelves, etc are more common ("indoor dangers") or one where durian fruits, falling tortoises, etc ("outdoor dangers") are more prevalent.

If the prophecy gets you either way, it's just a question of whether you'd rather meet your end indoors, perhaps while trying to reach something from a high shelf, with a deep sense of foreboding, or outdoors, probably in a completely unexpected fashion, or--shall we say?--completely out of the blue.

Actually, I think it was a Bearded Vulture. [0]

[0] https://en.wikipedia.org/wiki/Bearded_vulture

Today my iPhone suggested that I look at an automated album of home cooked food from my photos.

It was working pretty well as I scrolled through and then it came up with some poop on the ground that for some reason I had taken a photo of. Kind of disconcerting.

Your iPhone is just suggesting you to take some cooking classes, no big deal. It's its way of helping.

I mean it _was_ food.

talk about up an ante on eating your own dog food.. BE THE DOG.

I can't say I'm super surprised that their training data didn't include pictures of poop, as I imagine (or do I just hope?) that most people don't take many pictures of that.

Mmmm. Bowl of chocolate ice cream, tasty.

With the right auto-generated background music, this would have been a great dystopian 2020 summary.

The article’s title appears to be a reference to: https://en.m.wikipedia.org/wiki/The_Man_Who_Mistook_His_Wife...

That is a pretty obtuse reference to use without explanation.

It’s also an odd choice because it feels weird to call an AI a “who”. Despite AI replacing some human tasks, as with the camera operator here, the technology is not at all what we would consider a person.

Well, it's a very, very famous book.

Obtuse? It was even made into an opera, among other things.

https://www.latimes.com/entertainment/arts/la-xpm-2012-jun-1...

Not really. I got it instantly and I've never even read the book.

I understood the reference, but still felt irritated by the word "who".

Its a really famous book. And that particular story is about Visual Agnosia so its kind of apt.

When do we call it a "who"? Where is the line?

Sometimes I don't go down an aisle if I see the spill monitoring robot in my local grocery store. I think it's annoying, slow, dimwitted, and I can't communicate with it. Sometimes as a joke I'll even stare it in the camera and say "Your mother was a toaster!"

Where is the line where I'm no longer just a crazy person, but also a racist?

> That is a pretty obtuse reference to use without explanation. It’s also an odd choice because it feels weird to call an AI a “who”.

The slight incongruity of the word "who" is the explanation, or at least hint, that there is an unspoken reference in play.

So, if we assume that this game is important enough to somebody to televise, why would we not have volunteers to steer the camera? This does not look like a problem that needed AI applied to it, this has the look of somebody trying to apply AI to something that NI would have worked fine on.

Your basic soccer fan would have a pretty good idea of what part of the field is most interesting to be looking at right now, and would be happy to do it (for 10-15 minutes anyway) for free, and could do it remotely with a minimum of setup. You could have some season ticket holders (who can't use their tickets this year) volunteer to do it from their homes; they might even find it kind of exciting to do (for a while anyway), and you might quickly have a wait list.

A similar thing is done for nature-cams here: https://explore.org/livecams/brown-bears/brown-bear-salmon-c...

I don’t think you could trust a volunteer football fan to control this remotely.

Losing 3-0? Oops the cameras pointing at the sky for the last 30 minutes. Penalty to the opposing team? Oops I slipped and missed it. Football fans are petty and do absolutely ridiculous things.

You’d have to pay them something or find some other way to incentivise them, and you’re back to square one.

Then have a few “known good” people and then average their actions with the actions of everyone else doing it – so long as you do “virtual panning” on a cropped image to give the controllers sufficient feedback, that'll give you enough data to kick out trolls / misbehavers in a “Web of Trust” type manner, and even out most other issues that occur.

While I'm sure that would work, that sounds way more complicated than using computer vision to track the ball.

Unlike using computer vision to track the ball, it would probably work.

NI? what does it mean?

I am guessing “Natural Intelligence”, i.e. a human.

Eh, there's nothing new under the sun. The book A Robot Ping-Pong Player (https://mitpress.mit.edu/books/robot-ping-pong-player) contains an anecdote about the robot mistaking Marvin Minsky's head for a ping-pong ball... and taking a swing at him...

I worked for a small startup in copenhagen that worked on this exact problem. Our virtual camera solution was great, and detection of the ball worked most of the time, but white long socks would almost always take the camera's attention. Tracking a fast moving object at a distance is a hard problem indeed

In this case it seems someone could have asked the referee to wear a hat for the reminder of the game, but no one did.

This isn't just a "bad AI" problem, it's failure to plan for a "Plan B" in case automated ball tracking fails for any reason.

Some soccer teams have bald players. Should they also wear hats?

The AI should be fixed or abandoned in the long term, but it seems a single hat could have made this game much more bearable to watch.

Immediate fix now, better long term solution later.

Year 2060: "Granddad, when did they change the rules of football to say everyone on the field has to wear a dark head covering?"

Perhaps they can cut personal sponsorship deals, as they'll be getting so much more airtime in future

/jk

Well, it sounds like there aren’t a lot of support people for lower tier soccer... no camera operators, and no one even watching the camera feed.

Clothing in sports is highly regulated. For example, FIFA's Equipment Regulations are a hundred pages long[1].

Match officials in football (soccer, if you so will) very rarely wear hats (I can't remember seeing it even once). I'm surprised that I can't find any mention about headwear worn by match officials in the regulations, but regardless, they couldn't just don any old hat.

And as mentioned elsewhere in this thread, there are other things that also can trick up AI, like white socks. There are teams that might have had white socks for a hundred years, should they be forced to change their colours to make life easier for the AI?

[1] https://resources.fifa.com/image/upload/equipment-regulation...

Head coverings for referees would generally be a matter for national football associations as is the case with other parts of the uniform other than FIFA controlled international competitions. In the US, USSF lightened it’s stance on hats at lower levels because of skin and eye cancer cases involving “people everyone knew.” But you won’t see it a high level (e.g. professional matches) and young referees seeking to move up are discouraged from it in practice. And because the parallel system of High School and College officiating allow hats, USSF is probably more open than many national associations. The ADA also creates some legal liability as well that is unique to the US.

Does AI like this incorporate negative/positive punishment into learning as well as reinforcement? Is there an equivalent of slapping it and saying "No! Bad AI!" so it drops the association more quickly?

Yes but it should be avoided because the shock damages the hard disks' heads.

Everyone knows SSD is far superior :)

Would making the ball a "unique" colour work? (something that stands out, like magenta)

Earlier discussion here https://news.ycombinator.com/item?id=24955651

Mathematica mistook Stephen Wolfram for a plunger during an AI demonstration on the Lex Fridman podcast.

It passed the Turing Test.

Hey! He's a talented pure mathematician, and is absolutely not a plunger. To assert such is in no way social commentary.

Some time ago I needed ball video tracking for a demo at work. I failed to find any out of the box solution which worked satisfactorily. It was way easier to find solutions for seemingly more complex tasks of people tracking or human pose estimation, for which you have numerous publically available algorithms/models.

Balls dont have many unique attributes, and move a lot faster. Sonetimes humans are easier:)

Yes, yet we humans do it all right.

Either train more with bald players, or do better at choosing the 'thing that is being converged upon'. Humans would be able to follow a game regardless of the shape or color of the ball's appearance.

I imagine for soccer the task of determining where the ball is based on player locations would be a lot harder than you think.

'Spot the ball' was a staple of local newspaper competitions - an action shot of a game with the ball removed by presumably pre-digital photographic trickery. Entrants were invited to mark a cross where they thought the ball might have been, and the closest to the original centre co-ordinates won a prize.

So yeah, surprisingly tricky, but you could set up a big training set along similar lines.

I've seen that online with big costs for guesses and prizes. Many years ago but I'm sure it is still around. I thought about how to computationally determine the ball position until I read that the ball position was decided by an impartial referee's opinion.

Actually not the original co-ordinates, but a point arbitrarily decided by a panel. I suspect that's where most people went wrong!

https://www.theguardian.com/football/shortcuts/2015/jan/14/h...

It would be difficult from player 'positions' in a single image, but the example given is obviously wrong if you consider motion of player positions over time--those player positions going back and forth with the 'ball' moving by itself along the sidelines.

It's a hard task for young kids or anyone not used to watching sports. Ever watch a new sport with a child? They frequently lose sight/track of the ball/puck/etc.

A lot of it is learned heuristics. Things like how fast the ball normally travels, how hard it looks like the player threw or hit, game knowledge about places it makes sense for the ball to go, etc.

We don't train these algorithms while they are in active, live, use. If a human camera man mistook the bald head for the ball once in the game, he'd probably only do it once.

It reminds me of when I try to play any of the Smash Bros games. I can't keep track of my character on the screen ever. I don't know how people do it.

Those games are seemingly "visual cacophony to the point of difficulty and frustration" and I don't get the appeal. Most of my friends who enjoy it also tend to enjoy other fast-twitch games, as do I, but that game is a bridge too far.

You'd think at some point, if people were watching, they would've gotten the linesman to wear a hat or something.

Is this going to be the "interesting failure of AI" blog topic for the next year?

A new and worse case will soon present itself. (Though I think this one will be remembered.)

Maybe it's just in love.

Can't they just put a tracker inside the ball for trilateration?

lol i once played with imagenet and sending in Goku headshot (Super Saiyan ) it says it's a pineapple ! I mean I can see why !

Reply on news.ycombinator.com