February 27, 2016 at 21:11Google showed a neural network that could recognize a country from a photograph (as well as a city and street), even if the photograph was taken in a house
It’s easy to find out where the photo is taken if the Eiffel Tower, the Taj Mahal, St. Peter's Basilica, the Lincoln Memorial or Red Square are in the background. Developers from Google went further and made a neural network that could recognize a place from a photograph, even if it was taken indoors.
People can use not only the sights to understand where the photo is taken. The place can be determined by the dish in the restaurant, in the direction of traffic, by the cow on the street, the architecture of the buildings and the combination of all these factors. And what is the car capable of? PlaNet
Technology Developersdivided most of the land into 26,000 zones of various sizes, depending on the number of photographs taken in a particular area. Big cities got more “cells”, because more photos were taken in them, while in the countryside “cells” were larger. Seas, oceans, polar zones have passed.
A database of 126 million photos from the Internet was used along with their EXIF data. 91 million photographs were used to train the neural network, and the remaining 34 million were used to evaluate its work.
To test the effectiveness of the neural network, 2.3 million geo-targeted images from Flickr were used. 3.6% of images PlaNet recognized with accuracy to the street, 10% - with accuracy to the city. The country identified the neural network in 28.4% of cases, and the continent - in 48%.
This result was compared with the capabilities of a dozen travelers using the game GeoGuessr.com , in which you guess the place on Google Street View. PlaNet beat people with an average error of 1,131.7 kilometers. People were mistaken on average for 2320.75 kilometers.
According to one of the main researchers, Tobias Weyand, the advantage of the machine is that the neural network “saw” much more than any living person who has traveled around the world all his life.
The developers went further and began working with photographs taken on the premises. You can recognize them in cases where the photo is part of an album - the machine scans the albums completely and looks for the most specific images made in the same place
The neural network itself occupies only 377 megabytes.
People can use not only the sights to understand where the photo is taken. The place can be determined by the dish in the restaurant, in the direction of traffic, by the cow on the street, the architecture of the buildings and the combination of all these factors. And what is the car capable of? PlaNet
Technology Developersdivided most of the land into 26,000 zones of various sizes, depending on the number of photographs taken in a particular area. Big cities got more “cells”, because more photos were taken in them, while in the countryside “cells” were larger. Seas, oceans, polar zones have passed.
A database of 126 million photos from the Internet was used along with their EXIF data. 91 million photographs were used to train the neural network, and the remaining 34 million were used to evaluate its work.
To test the effectiveness of the neural network, 2.3 million geo-targeted images from Flickr were used. 3.6% of images PlaNet recognized with accuracy to the street, 10% - with accuracy to the city. The country identified the neural network in 28.4% of cases, and the continent - in 48%.
This result was compared with the capabilities of a dozen travelers using the game GeoGuessr.com , in which you guess the place on Google Street View. PlaNet beat people with an average error of 1,131.7 kilometers. People were mistaken on average for 2320.75 kilometers.
According to one of the main researchers, Tobias Weyand, the advantage of the machine is that the neural network “saw” much more than any living person who has traveled around the world all his life.
The developers went further and began working with photographs taken on the premises. You can recognize them in cases where the photo is part of an album - the machine scans the albums completely and looks for the most specific images made in the same place
The neural network itself occupies only 377 megabytes.