Look at the picture below and try to pinpoint where it was taken – you have 20 seconds and the file includes no GPS data (the answer is at the end of the article). In the absence of a recognizable element -like the Eiffel Tower, for example-you must admit that the task is arduous
well not for PlaNet, image geolocation recognition project Google unveiled by Technology Review. Without even really be reliable, this research program is already more efficient than humans. And he learns quickly.
Learn to read the photos
Led by a specialist computer programs dedicated to vision, Tobias Weyand, PlaNet program is a deep-learning software, ie he is not content to perform a task, he learns progressively and improves over time.
To learn things PlaNet, we had to give him the images to study. Oh, a straw: the basis of geo-localized image data assigned to this task no less than 126 million copies. Mr Weyand The team divided the database into a first segment 91 million photos for learning, and the rest was used for comparison results of its analysis.
Almost 50% success in the continent
Google teams have cut the earth in 26,000 areas whose size varies according to the density of shots – urban areas are smaller – and PlaNet started working. And he is facing 2.3 million geo-tagged images in Flickr. And the result is quite promising: the shot placement accuracy across continents PlaNet is already 48% – try to make the difference between the red earth of the Burkina Faso road that Australian bush
All young and already best we
the next step was that of confrontation with humans, 10 handpicked travelers.
And despite his young age, the program has won 28 of 50 GeoGuess innings game against humans, while proving more specific: when humans have an average 2320 kilometers precision geolocation, PlaNet twice more with only 1131 km. You can also play the game of “Where is that?” by visiting the geoguessr.com project page.
Coming soon to smartphone?
If it only worked on a supercomputer, PlaNet would be a research program as another. But according to its designer, the software takes up only 377 MB of memory and could thus entirely run on a smartphone!
If it does not give more details as to the need for access or not permanently in a database or the processor power required, it seems possible becomes your smartphone, in the coming years, both smarter and more relevant than you in analyzing the images. . What is also exciting … scary
Oh, I almost forgot: the photo of the beginning of the article was taken at the Sengi-neck in the state of Ladakh , Himalayan region of northern India. Not sure PlaNet is still developed for these regions …
No comments:
Post a Comment