Become a fan of Slashdot on Facebook


Forgot your password?
AI Google Privacy Technology News

Google Outlines AI-Based Number Reading For Street View Photos 68

mikejuk writes "A recent Google research paper outlines how it might use AI to read digits in natural images — specifically Street View photos. The idea is to automatically extract the number of each house as captured by Street View and then use this to improve the geocoding data returned by Google. When you next ask for directions to a particular address the new data could be used to show you a street view looking directly at the house you specified."
This discussion has been archived. No new comments can be posted.

Google Outlines AI-Based Number Reading For Street View Photos

Comments Filter:
  • by Joe_Dragon ( 2206452 ) on Sunday December 18, 2011 @12:35PM (#38417036)

    does it crash?

  • by Anonymous Coward on Sunday December 18, 2011 @12:35PM (#38417042)

    I've poked around the US Census data which has geographical coordinates of pretty much everything you can imagine - streets, natural landmarks, including addresses.

    How does most mapping software get its dataset? Why does Google need to find street numbers from the photos? (Probably because a public dataset like this isn't available globally) I imagine that Census data is a free alternative but professional geolocation data requires big fees.

    • by djl4570 ( 801529 )
      An underlying reason for this is to better correlate the map with the objects placed on the map. Accurately place the Pita Place on the map between the Jamba Juice and the Charbucks instead of putting it on the other side of the parking lot where there isn't a building.
    • by skids ( 119237 )

      At least some of the current store of street number data came from publishing HITS on Amazon Turk, so it's wetware-derived.

    • A lot of this comes from the R.L. Polk database. Those guys hire out at minimum wage. Would you trust that data? I know a few guys that sat at a Denny's for a week just reusing info from a phone book for their daily logs. This will by a lot more accurate, since I don't know too many people that put the wrong street number on their houses.
    • Why does Google need to find street numbers from the photos?

      Google doesn't NEED to find the street numbers from photos. But evidence suggests that the data Google currently uses isn't quite accurate. Google thinks it can improve on that accuracy, for almost no additional cost.

  • Perhaps I'm not thinking hard enough, but this seems to me to be an development from one of the $InternetSocialMediaOverlords that doesn't seem creepy.

    I mean, this is a nice feature that will save time and be useful; but it doesn't go revealing yet more personal information, or infact anything that you couldn't do yourself by browsing streetview a bit.

    So yeah, a new shiny that doesn't yet make things worse or closer to 1984.

    • by Surt ( 22457 )

      This improves the ability of the gestapo to avoid getting lost on their way to pick you up.

    • This isn't a development of the $InternetSocialMediaOverlords. It's just a bunch of nerds who know some algorithms and enjoy solving practical problems who thought this would be fun to attempt, and happen to work for the $InternetSocialMediaOverlords. Trying to look for a conspiracy in this would be akin to looking for a conspiracy in the way the leaves pile up in your driveway in the fall.

    • That's right, this isn't creepy, because StreetView itself is the creepy part. This is just refining it.

  • Captcha just failed (Score:2, Interesting)

    by Anonymous Coward

    If it can read a digit in a pattern then capchta just became useless as a way of keeping bots out of a website

    • by djl4570 ( 801529 ) on Sunday December 18, 2011 @12:46PM (#38417126) Journal
      Then we'll have to migrate to "Cutest kitten"
    • by jdpars ( 1480913 )
      Except that house numbers are generally made to be readable at a distance, whereas captchas are made to be readable up close, with scrutiny.
      • Captchas are made to be unradable. The average human success rate at some Captchas is far lower than that of a OCR-based cracker.
    • You should look up how the captcha system works (or reCAPTCHA, anyways). It is digitally scanned old print (such as old editions of the New York Times). They feed it through several OCR systems, and the ones that it doesn't work on get thrown into the CAPTCHA system for humans to identify. This allows them to a) digitize massive amounts of old print material (using the humans interpretation of words that can't be read by computer) and b) ensure large sources of CAPTCHAs which are unreadable by present OCR s

      • by satuon ( 1822492 )

        They can't use the words from old print because they don't know what it reads, so they can't know if you've answered correctly. The captcha that is used for verification is computer generated, that's why reCAPTCHA has 2 words - one is generated (hence they know what it is), and one is from old print. If you answer the computer generated captcha it's assumed that you've answered the other one as well. They also send the same old print word to 2-3 people to see if their answers match. That way Google gets a c

        • you are unfamiliar with the tech underlying recatpcha []

          why do you think it requests two words?
          one word it has classified one it hasn't.
          If you input the classified word correctly you pass. The input of the unclassified word is buffered.
          Unclassified words that get consistent results are added to the classified word pool and used as validators.

          It is a basic ml application whose computing is outsourced to the humans on the Internet who want to post
          on and a very elegant solution as well. if you wan

      • by Ihmhi ( 1206036 )

        Being pedantic here, but one of the words was probably read by OCR. One word in reCAPTCHA is unknown and the other is verified (but blurred, struck out, color inverted, etc.). It's likely that the machine-readable word was originally caught properly by OCR.

  • It's clear now.

  • by Animats ( 122034 ) on Sunday December 18, 2011 @12:47PM (#38417134) Homepage

    I've been wondering when they'd make that work.

    In retail areas, street numbers tend not to be too prominent. It may be necessary to read business signs and use that data to disambiguate addresses. This would help to clean up the phony-business problem in Google Places. An alternative is to use real estate records, as the USC Geocoder [] does for some areas, to get a solid lock on address vs. physical position. But that data is only available for some areas. There are also the Census Bureau's TIGER/LINE [] files, but they're US only and not complete for the entire US.

    Outside the US, this is likely to be more useful. If you have a few street numbers and a few business signs per block, you can infer the rest reasonably accurately.

    • In retail areas, street numbers tend not to be too prominent.

      I have found that to be true and don't understand it. While businesses loudly proclaim their names and logos, their street numbers are often barely visible if present at all. When dealing with heavy traffic, you often can't just leisurely slow down and rubberneck to search for some street number clue in an unfamiliar area. (Well, pre-GPS anyway.) More than once I've just given up trying to find a business while driving and gone

  • Is this the paper Peter Norvig (quite tangentially) referred to near the end of this article [] in the MIT Tech Review? Can anybody confirm?


  • If the proliferation of CCTV using facial recognition could be tied into Google allowing you to Google for faces and where they were last seen.

    So using a 3 Dimensional recognition algorithm and saving that data plus the location would then allow you to drag a picture of your friends face into Google (as you already can) then get results of where that face was last seen.

  • Could humans, with their supremely evolved optic nerve/brain, outperform any A.I. that Google might have now or in the near future? Might it even be cheaper? (Just how much does difficult A.I. cost in the cloud anyway?). Maybe even if A.I. is much cheaper, it would still be useful for using humans for the difficult cases, or as error correction. Best would be for the humans to train the A.I. (and themselves out of a job!)

    I always thought this was the only possible reason how the machine intelligences in

    • I'm sure that it could be done by the Mechanical Turk. But considering the amount of data that they've got ... possibly it would be too expensive. The Turk may be flexible and relatively cheap, but it's still got a finite cost per computation, and meat-puppets are not particularly cheap. (It might be useful for cross-checking hit rates etc though. Quicker and cheaper than coding and testing a second/ third/ fourth algorithm in detail.)
  • by doudou42 ( 691076 ) on Sunday December 18, 2011 @01:28PM (#38417440)

    It's interesting to note that one of the co author is also the teacher for stanford free classe on Machine Learning and that the last lesson of the course was on this topic...

    • You noticed that to? It's nice of him to pass along some of the stuff he is working on as a lesson for the class.

      • Don't most profs do that? Isn't that the whole point of having PhDs doing both research and teaching? I know most of my 4th year profs talked about what they were working on in class...
      • nice of him to pass along some of the stuff he is working on as a lesson for the class.

        Almost as immoral as using the Mechanical Turk! (As someone else suggested upthread.)

  • Will it work with those annoying recaptcha boxes?
  • It's machine learning.

    There's a difference.

    • It's machine learning.

      There's a difference.

      Please explain the difference, then. At Stanford, UBC, MIT (I assume other places as well), their artificial intelligence labs include their machine learning groups.

      For example, at Stanford, the Artificial Intelligence Laboratory includes the research groups of Daphne Koller, Andrew Ng, and Sebastien Thrun, who all describe their research as Machine Learning.

      Andrew Ng, a machine learning research and instructor for the online machine learning course ( is direc

  • they were trying to appeal to english speaking olympics visitors in 2008, and the translation server crashed, but they thought the error message was the translation, so we get: []

    so would it be like divide by zero if machines try to map the real world and encounter a bit of the real world mimicking machine world fail?

  • by MrL0G1C ( 867445 )
    What the f*** would the point of that be? in the UK they blur out door numbers and street signs which is bloody inconvenient when you're researching visiting a place.
  • You open a shop in the same road as a very sucessful one and put their number sign (and the shop name, in case they use OCR on it) just when the google car passes by......

  • by Mister Liberty ( 769145 ) on Sunday December 18, 2011 @02:15PM (#38417784)


  • I remember when I first heard that the spy satellites could read a cars license plates.

    I thought that's good trick, not only do they have to view small characters they
    had to do it sideways.

  • In Saskatchewan, we have a Crown Corporation that is responsible for maintaining, updating, and distributing that map data for the province. It's accurate data -- collected using actual transits and GPS systems. Google can not improve on the data quality by reading street signs with AI at all.

    I suspect it has a lot more to do with licensing issues for Google wanting to republish the data to the general public, something that I'm sure our crown corp would frown upon as it would cut out their geodata sal

  • ... and get an army of volunteers to help them.

  • Most StreetView images are soooooo low res that you can't make out street numbers on most houses.

    Maybe businesses with large street number signs, but not houses.

You can measure a programmer's perspective by noting his attitude on the continuing viability of FORTRAN. -- Alan Perlis