Slashdot is powered by your submissions, so send in your scoop

 


Forgot your password?
Close
typodupeerror
×
Google AI Earth News Science Technology

Google Unveils Neural Network With Ability To Determine Location of Any Image (technologyreview.com) 116

Posted by BeauHD from the neural-networking dept.
schwit1 writes: Here's a tricky task. Pick a photograph from the web at random. Now try to work out where it was taken using only the image itself. If the image shows a famous building or landmark, such as the Eiffel Tower or Niagara Falls, the task is straightforward. But the job becomes significantly harder when the image lacks specific location cues or is taken indoors or shows a pet or food or some other detail. Nevertheless, humans are surprisingly good at this task. To help, they bring to bear all kinds of knowledge about the world such as the type and language of signs on display, the types of vegetation, architectural styles, the direction of traffic, and so on. Humans spend a lifetime picking up these kinds of geolocation cues. So it's easy to think that machines would struggle with this task. And indeed, they have. Today, that changes thanks to the work of Tobias Weyand, a computer vision specialist at Google, and a couple of pals. These guys have trained a deep-learning machine to work out the location of almost any photo using only the pixels it contains.
This discussion has been archived. No new comments can be posted.

Google Unveils Neural Network With Ability To Determine Location of Any Image More

Google Unveils Neural Network With Ability To Determine Location of Any Image

Comments Filter:

  • Can it figure out where goatse was taken? (Score:4, Insightful)

    by Anonymous Coward on Saturday February 27, 2016 @06:33PM (#51600943)

    Can it figure out where the notorious goatse photo was taken?

  • Indoor photo location (Score:1)

    by Anonymous Coward
    I'm well traveled enough to recognize certain kinds of indoor fixtures as being typical for specific countries, or at least regions. One giveaway are electrical sockets. You can learn a lot from the details on those. Window shapes, frames and latches also reveal a lot.

  • Almost all = not that many (Score:3)

    by gweihir ( 88907 ) on Saturday February 27, 2016 @06:38PM (#51600965)

    In mathematics, "almost all" means "all, except for a finite number". As the number of pictures in this universe is finite, this could even mean this "magic" machine cannot do it at all. I guess the headline was written in this spirit.

    • In mathematics, "almost all" means "all, except for a finite number". As the number of pictures in this universe is finite, this could even mean this "magic" machine cannot do it at all. I guess the headline was written in this spirit.

      I'm not sure the author was using the phrase in a mathematical sense. More likely it was in the probabilistic sense of a very high percentage of successes in a number of finite trials.

      BTW, "almost all" in mathematics (specifically, real analysis) can also mean "all, except for a set of zero measure." Consider for example, the Cantor set [wikipedia.org] which has an uncountably infinite number of members lying in the interval [0,1], yet it has measure 0. So, if you construct a function f(x) on [0,1] that has the value 1 at

      • Re: (Score:2)

        by Dahamma ( 304068 )

        I'm not sure the author was using the phrase in a mathematical sense. More likely it was in the probabilistic sense of a very high percentage of successes in a number of finite trials.

        Not really... the author was mostly using it in the Internet clickbait sense.

        If you try it out, it's basically guessing the location of a Google street view image on Google maps. Hardly "determining the location of any image".

        Still, it is a lot of fun. Somehow I scored > 10000 points on the "Paris" map even though I was there my first (and only) time last year. Guess it helps to have walked the shit out of a city...

      • Re: (Score:2)

        by gweihir ( 88907 )

        I am very sure he was not. I was though.

    • In mathematics, "almost all" means "all, except for a finite number".

      Does it always mean that, or almost always mean that?

    • This is very true, the number of pictures out there is actually very large though. Mid 2014, we humans were posting 1.8 BILLION pictures a day on Internet. I guess it's doubled since then. Most of the data is now created automatically by Internet of Things connected objects (your Fitbit, etc) and I guess the same trend will apply to photography. Makes sense to have automated processes / tools to help us make sense of all this nonsense. The tricky part is, as always with tech, how will they protect our priv

    • In mathematics, "almost all" means "all, except for a finite number".

      It could easily be an infite number as well, especially if the "all" is uncountable in the first place.

    • In mathematics, "almost all" means "all, except for a finite number"

      Hmmm ... did "mostly kinda sorta" become a rigorous mathematical construct when I wasn't looking?

      As the number of pictures in this universe is finite

      While waiting for a seat in a restaurant last night, I watched 3 teenage girls taking selfies. I assure you, this is not true.

      And, as a bit of a photography junkie who does a lot of macro photographs and weird things which aren't just straight representational and therefore often don't have m

  • It can't possibly determine the location of any image. Many will simply have too little location data to form a reasonable guess.

    • It can't possibly determine the location of any image. Many will simply have too little location data to form a reasonable guess.

      Precisely. That's why TFS and TFA qualify it by saying it can determine the location of almost any image.

      • Re: (Score:1)

        by Dahamma ( 304068 )

        Which is also totally untrue. It should have said "it can determine the location of almost any image ON GOOGLE STREET VIEW".

        • But that's not really a neural network then. To "find" subsets of an image within a set of other images, even a (large) database of them is relatively trivial (from a coding perspective) even if things like color, lighting and compression change and doesn't require any "special sauce" like AI or neural networks.

          What's next: Apple Siri can guess songs it hears from their iTunes catalog? FaceBook can identify faces within pictures? YouTube has a search function.

          • Re: (Score:2)

            by Dahamma ( 304068 )

            But that's not really a neural network then

            Wha? It's either using a neural network or it isn't. Whether it's a neural network or not doing the computing has nothing to do with the results or the efficiency of the search vs more traditional computing methods.

            And I wasn't commenting on the methods or even the effectiveness of the program, just the (probably clueless non-tech) editor's assertion that it applied to "almost any image".

            • But that's not really a neural network then

              Wha? It's either using a neural network or it isn't. Whether it's a neural network or not doing the computing has nothing to do with the results or the efficiency of the search vs more traditional computing methods.

              If the neural network hasn't been trained in some way on the images it sees, then it's not fair to ask it to recognize them.

              The same criterion applies to a human being.

          • To "find" subsets of an image within a set of other images, even a (large) database of them is relatively trivial (from a coding perspective) even if things like color[...]
            Then sketch us your algorithm here and farm in your Nobel Prize.

            • Re: (Score:2)

              by guruevi ( 827432 )

              How the hell do you think face (or any sort of image) recognition works? Never heard of license plate scanners? Those things are relatively trivial.

              • And? You seem not to know the difference in trivia between a license plate scanner and your "simple proposal" of simple picture matching ;D

                How the hell do you think face (or any sort of image) recognition works? 99% of it works by "knowing what you are looking for and figuring where it is in the picture" ...

                Those things are relatively trivial. No, they are not. There are dozens or 100ds of algorithms involved to look for a single feature. I doubt you know any of them.

    • On the other hand, it could settle quite a few questions. Give it shots of the moon landings or the alien autopsy.

    • I can tell with 99.9999999999% accuracy where ANY photo was taken, without even looking at the pixels:

      Earth.

      What the article fails to mention is any reasonable measure of geographical range. It's one thing to say, 'this photo was taken in China'. It's quite another to say, 'this photo of cat food was taken in Bob Smith's kitchen at 22 Kings Mews, London, SW1'.

      The real news here is the claim that the software/network can pinpoint as well as a human can. And I'd like to see that tested.

      • What the article fails to mention is any reasonable measure of geographical range.

        "PlaNet is able to localize 3.6 percent of the images at street-level accuracy and 10.1 percent at city-level accuracy," say Weyand and co. Whatâ(TM)s more, the machine determines the country of origin in a further 28.4 percent of the photos and the continent in 48.0 percent of them.

        The real news here is the claim that the software/network can pinpoint as well as a human can. And I'd like to see that tested.

        "In total, PlaNet won 28 of the 50 rounds with a median localization error of 1131.7 km, while the median human localization error was 2320.75 km."

        I know it's standard practice here to comment without reading the articl

        • Em... right, then. You are correct, I did not RTFA.

          I meant to write "post", not "article".

          But that does not excuse my laziness or misspeaking. My apologies.

    • Considering that plenty of pictures will be "Golden Gate" and "Eiffel Tower" I would assume you would be surprised how many places you actually can guess.

      Don't have to been there, movies are enough.

  • Try it yourself (Score:5, Interesting)

    by mobby_6kl ( 668092 ) on Saturday February 27, 2016 @06:41PM (#51600989)

    Well the "Guess the location" thing, not the NN :)

    There's a site that basically opens StreetView at random around the world and asks you to place it on the world map. As the summary explains, you can use a number of clues to generally place photos surprisingly accurately. Used to play this occasionally at work, we really liked that it challenged you to think about all these things that you know about the countries and regions around the world.

    https://geoguessr.com/ [geoguessr.com]

    • Re: (Score:2)

      by adolf ( 21054 )

      As the summary explains, you can use a number of clues to generally place photos surprisingly accurately.

      Feedlot? John Deer tractor mowing? White lines instead of yellow? A squished rabbit on the road?

      (I wasn't sure until I saw the rabbit.)

      • Feedlot? John Deer tractor mowing? White lines instead of yellow? A squished rabbit on the road?

        (I wasn't sure until I saw the rabbit.)

        If you're the one who hit that rabbit you could probably place the image pretty damn accurately!

        Generally, if it looks like America but with fewer trucks and everyone drives on the opposite side, it's Australia. If looks quite like America but something's slightly off, that's Canada. If everyone drives on the opposite side and it's cloudy, wet, and depressing, thats's the UK. If it just looks depressing, that's Russia.

        Here's [slashdot.org] my result from the fist try after not playing for a while. My biggest challenge see

        • Re: (Score:2)

          by adolf ( 21054 )

          No vehicles in sight, except for the mowing tractor. No road signs nearby, either.

          It was Australia. Rabbits are ridiculously common there.

          It's an odd game, it is habit-forming, and it gets easier as it gets played... It should be part of a high school geography or social studies class.

          And there goes my afternoon.

    • Re: (Score:2)

      by MrL0G1C ( 867445 )

      I just hunted around for street signs and googled them! Cheating no doubt but fiendishly difficult otherwise when you've got nothing to go on other than a road and some fir trees.

      • That's not much fun though, is it :)

        If you have realistic expectations, often you don't really need much more than a road and some trees - you can tell the country by the road signs (like, speed limits and stuff, not city names :D) and lane markings and the region by the trees. But sometimes you're unlucky and get a dirt road in the middle of a pine forest or something, then you're screwed.

        One place I would expect the AI to do a better job is identifying cities. I can ID NYC, DC, Paris or Berlin, but I can'

        • Re: (Score:2)

          by MrL0G1C ( 867445 )

          Actually it's still a challenge, I had to type in Russian letters, zoom in on a tv screen, hunt for road signs, still make guesses based on countryside, road markings etc. My places were more diverse, I got Argentina middle of nowhere, ditto for Iceland, Russia and Canada.

        • I can't tell one anonymous shithole from another

          You're on slashdot.

          No, no need to thank me. I'm into public service.

    • Per the original paper [arxiv.org], that's what they use to train the NN and as a goodness metric.

    • Re: (Score:2)

      by Hentes ( 2461350 )

      There are many kinds of random. The site chooses a photo without taking into account where it was taken, and because the density of photos varies greatly with location, even a simple statistical algorithm using no data from the images could get fairly accurate at guessing.

    • Thank you. I had run out of things to waste time on the internet.

  • Almost any photo = 10.1% (Score:5, Informative)

    by ebcdic ( 39948 ) on Saturday February 27, 2016 @06:46PM (#51601009)

    According to the article, it can identify 10.1% of the Flickr images it was tested on at "city-level" accuracy.

    • Re: (Score:1)

      by Anonymous Coward

      NOBODY reads the articles. The brilliant minds on Y Combinator news also only read the headline.

    • Compared to a human that's pretty dam good.

      • Compared to a human that's pretty dam good.

        Depends on what you're trying to measure.

        Computers are always faster and better at dealing with large quantities of data. For example, if I want to search a book-length text, my computer can instantly find all strings matching a given pattern. Obviously, a human could do this too, but it would take a while, and if rushed, the human might miss some.

        But I don't think we'd say that a simple search for a string indicates "intelligence" in the AI sense, no matter how much faster or more accurate a computer

  • Uh oh (Score:3)

    by rampant mac ( 561036 ) on Saturday February 27, 2016 @07:13PM (#51601103)

    Guess us old-timers will finally find out where the goatse guy is hiding out. :(

    • Re: (Score:2)

      by OzPeter ( 195038 )

      Guess us old-timers will finally find out where the goatse guy is hiding out. :(

      He's probably sharing a house with tub girl

      • Guess us old-timers will finally find out where the goatse guy is hiding out. :(

        He's probably sharing a house with tub girl

        Sorry, but it's the Lemon Party dude. Now go wash your eyes out with bleach.

    • Duh, goatse guy doesn't have a mailing address, he has a PO box. In his ass.

      You can mail him, but you can't female him.

      He's not hiding out, his ass is everywhere.

      Last time someone said, "Up yours," he said, "You'll need at least two."

      And if A.I. ever needs to be put down, and a paradox fails, "Where is goatse guy" should give you a few days to escape while a computer tries to figure out "Wtf is this," "I can't even," and finally, "How many jelly bellies worth of excavated arsehole is that, approximately?"

    • ...in a maze of twisty little passages, all alike.

  • Take a shot if.... (Score:1)

    by Anonymous Coward
    you watch a movie with it and it returns Vancouver

  • Now, where does it say the moon landings were filmed?

  • ... my car keys.

    • ... my car keys.

      They're in your car, and it's moving at exactly 88 mph. Not even Google can beat physics.

  • How many of the human contestants had been trained using 91 million tagged images specifically for this purpose?

  • How will it cope with an indoor picture from Australia with an Eiffel tower picture hanging on the wall?

    Any human will conclude the Eiffel tower here does not mean the picture is from Paris, but a computer?

  • Artists can sense location by the quality of light (Score:5, Interesting)

    by Daniel Matthews ( 4112743 ) on Saturday February 27, 2016 @10:27PM (#51601657)
    Perhaps not all of them but it is a thing many talk about, that different locations have a distinct quality to the natural light there. Not just location either, but seasonal variation too. It would not surprise me if the NN was also sensitive to these clues. Which makes me wonder can Weyand el al get their NN to tell them additional things about the image such as time of day and day of year, perhaps even in some cases actual year because it can cross reference images in the same location with known dates against an image that it has geo-located and know what the weather (lighting conditions) were for that point in time near that location.
  • This is why I only use organic pixels for my photos. None of those unnatural digital markers contaminating my files!

  • I wonder if they're planning on using this to track criminals such as child pornography creators etc online. It would certainly be a good use for this. I doubt it's accurate enough to capture a house etc (unless somebody was dumb enough to have other pics in the same house/room), but it might help regionalize it.

  • Google claims to have this so-called 'deep-learning machine' can identify the location of any photo? I will bet Google 1 BILLION dollars that they can't identify the locations of most of the photos I have...all pre 90s photos taken all over the US with absolutely NO landmarks in the background or other key deciding factors. Someone at Google got waaay too cocky about their 'slightly improved' algorithm. On the other hand, I once knew a guy who invented a machine that could generate 400% of the power requ

  • The results make for interesting reading. To measure the accuracy of their machine, they fed it 2.3 million geotagged images from Flickr to see whether it could correctly determine their location. "PlaNet is able to localize 3.6 percent of the images at street-level accuracy and 10.1 percent at city-level accuracy," say Weyand and co. Whatâ(TM)s more, the machine determines the country of origin in a further 28.4 percent of the photos and the continent in 48.0 percent of them.

    Oh,. boy, 3.6 percent. Ye

  • Okay Google... Tell me exactly where on the planet THIS PHOTO [staticflickr.com] was taken.

    I'll wait...

Slashdot Top Deals

C'est magnifique, mais ce n'est pas l'Informatique. -- Bosquet [on seeing the IBM 4341]

Close