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.
Paper: http://arxiv.org/abs/1602.05314
Submitted via IRC for Bytram
(Score: 3, Interesting) by Beige on Saturday February 27 2016, @07:06PM
It would be interesting to know how well the "superhuman neural network" would perform for a subset of photos of locations which don't exist in the 91 million photo dataset? It's an interesting experiment, but on the face of it 10% accuracy for identifying the correct city and 3% accuracy for street-level accurate identification isn't particularly useful as such.
However, this could become quite interesting if they e.g. trained the network against all of Google's Street View data, as this could potentially become an alternative method for e.g. retroactive geotagging of image data without a GPS tag. And yes, I know people will say a network of such a size is years away due to the scale, but here the 28.4% accuracy for determining the country of origin of the image with a "cheap" neural network is promising as it'll allow for splitting down and optimising the search (i.e. similar to the old quadtree algorithm but for neural networks).
(Score: 0) by Anonymous Coward on Saturday February 27 2016, @08:38PM
Can it beat ReCaptcha? If it can be trained to "select all images with street signs" it could make web-browsing far more pleasant.
(Score: 2) by takyon on Saturday February 27 2016, @09:36PM
Can it (or other neural network technique) beat ReCaptcha at a low cost? There is a financial incentive to beat the captchas.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Interesting) by JoeMerchant on Saturday February 27 2016, @09:02PM
Where in the world becomes a whole lot easier when there's a GPS coordinate encoded in the image.
🌻🌻 [google.com]
(Score: 2) by frojack on Saturday February 27 2016, @09:36PM
Yup. GPS is a requirement on all my cameras nowdays. Won't even consider one without it.
I had too many interesting photos taken on trips that I could not even locate knowing the sequence and date they were shot. Especially road trips, but also some multi-stop-over air travel.
Google Goggles app could sometimes locate landmarks for me, but its reliability has fallen dramatically over time.
Some even contain a compass so you can shoot a landscape, and not only map out the locations of each shot, but also the direction, the path you took through the landscape, etc. I'm like photographing ruins, old earthworks, battle scenes, historical sites etc.
Cameras in just about every price range include GPS these days.
http://snapsort.com/explore/best-digital-cameras/have-a-gps-24-months-recent? [snapsort.com]
No, you are mistaken. I've always had this sig.
(Score: 3, Insightful) by Jiro on Sunday February 28 2016, @11:53PM
This headline isn't as clickbaity as the Slashdot headline (which asserted that it can find the location of "any image" when it could find 3.6% of them at street level accuracy), but this still isn't as impressive as it sounds. "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". The program and the humans are tying for last place, and there are no other places.