Bashny.Net
Google showed the neural network that is capable to learn to photograph (as well as the city and the street), even if the photo is made in house
Bashny.Net
It is easy to find out where the picture is taken, if the background - the Eiffel Tower, Taj Mahal, St Peter's, the Lincoln Memorial and Red Square. The developers of Google went ahead and made a neural network capable of check the location on the picture, even if it is made in the room.
Users can use not only the sights to see where the picture was taken. Location can be determined by the dish in a restaurant, in the direction of the traffic, a cow on the street, the architecture of the buildings and the reconciliation of all these factors. A machine that is capable of?
PlaNet technology developers shared most of the land to 26,000 zones of varying size depending on the number of photographs taken in a particular area. Big cities have gained more "cells", as they have done more photos, while in rural areas "cells" were larger. Seas, oceans, polar areas missed.
It was used base of 126 million photos from the Internet together with their EXIF-data. 91 million photos were used for training the neural network, and the remaining 34 million - for the evaluation of its work.
To test the effectiveness of using a neural network 2, 3 million geotargetirovannyh images from Flickr. 3, 6% PlaNet images learned up to the street, 10% - up to the city. Land of the neural network has identified 28, 4% of cases, and the continent - 48%.
This result is compared with the possibilities of a dozen travelers with games GeoGuessr.com, in which you guess the location on Google Street View. PlaNet beat people with an average result of an error in 1131, 7 kilometers. People were mistaken on average in 2320, 75 kilometers away.
According to one of the main researchers Ueyanda Tobias (Tobias Weyand) Machine advantage is that the neural network "seen" so much more than any living person, who traveled around the world for life.
The developers have gone further and began to work with photographs, which are made on the premises. Learn they can be in cases when a photo is part of the album - the machine scans the full albums and looking for the most specific images taken in the same place
Neural network itself is just 377 MB.
Source: geektimes.ru/post/271738/
Users can use not only the sights to see where the picture was taken. Location can be determined by the dish in a restaurant, in the direction of the traffic, a cow on the street, the architecture of the buildings and the reconciliation of all these factors. A machine that is capable of?
PlaNet technology developers shared most of the land to 26,000 zones of varying size depending on the number of photographs taken in a particular area. Big cities have gained more "cells", as they have done more photos, while in rural areas "cells" were larger. Seas, oceans, polar areas missed.
It was used base of 126 million photos from the Internet together with their EXIF-data. 91 million photos were used for training the neural network, and the remaining 34 million - for the evaluation of its work.
To test the effectiveness of using a neural network 2, 3 million geotargetirovannyh images from Flickr. 3, 6% PlaNet images learned up to the street, 10% - up to the city. Land of the neural network has identified 28, 4% of cases, and the continent - 48%.
This result is compared with the possibilities of a dozen travelers with games GeoGuessr.com, in which you guess the location on Google Street View. PlaNet beat people with an average result of an error in 1131, 7 kilometers. People were mistaken on average in 2320, 75 kilometers away.
According to one of the main researchers Ueyanda Tobias (Tobias Weyand) Machine advantage is that the neural network "seen" so much more than any living person, who traveled around the world for life.
The developers have gone further and began to work with photographs, which are made on the premises. Learn they can be in cases when a photo is part of the album - the machine scans the full albums and looking for the most specific images taken in the same place
Neural network itself is just 377 MB.
Source: geektimes.ru/post/271738/
Tags
See also
Dome on Red Square (5 photos)
Red Square three hours before the New Year (9 photos)
Long exposure
10 facts about the sculptor and architect
The main thing is not the photo and background! (2 photos)
Predictions for 2014
Unexpected predictions for 2014
Cathedral of the Holy Family (La Sagrada Familial) (51 photos + panorama)
The smallest of the country (10 photos)
