Online algorithmic event detection from realtime mobile photos

http://detectd.teleportd.com

by spolu • 14 years ago

81 21 14 years ago

detectd.teleportd.com

gstar• 14 years ago

This is really incredible - I first heard of this when the Tottenham Court Rd incident happened a few weeks ago in London, and teleportd had images and detected the event before Sky News.

Well done guys! Really cool to see it launched.

spolu6• 14 years ago

Thanks! You should check out the various commencements in NYC going on right now :)

smackfu• 14 years ago

Now it just needs to learn that "commencement" and "graduation" are synonyms.

matthieurouif• 14 years ago

I discovered pictures of this car entering the French metro through teleportd. They had way more images than news website. I also love to teleport myself to Rio and San Francisco where part of my family is.

dclowd9901• 14 years ago

This is such a great idea. And imagine if first-hand reporting could be confirmed using a tool like this.

There are so many interesting and potential uses from such a product.

spolu6• 14 years ago

Thanks! we're still tweaking our algorithm to lower the number of false-positives, but some news outlets already said they would indeed be interested by having this kind of stuff!

leonbuchard• 14 years ago

Is that Bono on the 4th photo ?

gabhubert• 14 years ago

That or a guy doing a seriously good face expression.

mcdillon• 14 years ago

This is pretty awesome and its a great way to spark off additional analysis, such as triggering precise twitter feed filters for a specific topic.

ragmondo• 14 years ago

Isn't this what "Color" was supposed to do (amongst many other things that is) ?

spolu6• 14 years ago

They probably said that once :)

loceng• 14 years ago

Please don't sell to Facebook. You'll have more value being a de-centralized company.

spolu6• 14 years ago

Not on our TODO list yet!

rberrehili• 14 years ago

The power of technology! A laptop, some (a lot of) lines of code and here you are!

Congrats guys! ;)

maxjaderberg2• 14 years ago

really cool tech here! have you guys published any papers on this?

spolu6• 14 years ago

Not yet but actually we're working on it. We hope to publish something before end of 2012! Any idea on the good confs for such subject?

maxjaderberg2• 14 years ago

Looking forward to it! Not sure for your subject, as I am computer vision guy. In fact, I work on large scale image recognition and there is something about this which I love...there's a lot of potential for image analysis here.

ljd• 14 years ago

Congrats, in my humble opinion this looks like a game changing idea. I don't see any reason why this can't become a legitimate news source or the visual version of reddit where upvotes come from density ratings.

I'm pretty sure you've thought of this but if not here is some algorithm advice: you should start recording baseline density referenced by time (maybe use Fourier to detect time patterns?) so you can start filtering out the noise. That way you can filter out the few thousand pictures of the Empire State Building everyday unless something happens at the ESB in which case density would dramatically increase relative to time.

Also, this would be great to see visually on a map where when you zoom farther and farther in the clusters get redrawn so I can see micro events happening within the major event.

Great job!

ziends• 14 years ago

great job

WiseWeasel• 14 years ago

I'm curious how things like "bootylicious" in San Francisco and what appear to be mostly personal photos like "griffith observatory" in Los Angeles, and perhaps artistic things like "lighting shadow" in San Francisco made the cut. Is it that at least two different accounts posted photos with those tags in the same general vicinity within a certain time elapsed?

Maybe it would help filter out less relevant results if events without any photos featuring large groups or crowds of people are removed, since those typically go hand-in-hand with notable events. If there could be a way to separate out sporting and music events, that would be nice as well.

Prime featured content would be large groups of people in the streets, with bonus points if there are things on fire. Another would be natural disasters (floods, tornadoes, hurricanes, earthquakes, tsunamis), which most likely do not contain pictures of crowds in the streets, but probably do contain damaged or flooded buildings. Another might be a disease outbreak, perhaps with pictures of individuals with lesions or other visible symptoms, maybe in a medical setting. War zone events like the shelling of a neighborhood somewhere would be interesting, and might fall under the damaged building filter for natural disasters. Maybe pictures of tanks or other heavy military equipment rolling into a town, or helicopters flying overhead. UFO sightings. Pictures of crashes and accidents are always popular as well, and might only have a few emergency personnel around. Political events would probably fall under the large crowd filter. What else?

Online algorithmic event detection from realtime mobile photos

Related Stories