RSS feeds in the directory: 374
Added today: 0
Added yesterday: 0
We have released a new game, you can play with Alice, — "Guess the city in photography". The game mechanics are simple: you are shown a random picture from Yandex.Cards, and you guess which of the four cities of Russia it is made. The game itself is quite complicated. In the pictures catches all in a row — and the most significant sights and places are in every town: buildings, churches, shopping centers, and residential districts. To know Moscow's red square or Kazan mosque Kul Sharif is not difficult. But in order to determine what city of Russia lifted a nine-panel, you have to be a real historian.
the Correct answer supporting details: terrain, plants, vehicles, banners and signs. If the photo you see a Lombardy poplar, it is unlikely that it was taken somewhere in the North. If the picture "Stalin", it's definitely not Noyabrsk and New Urengoy — they are based in the 1970-ies. The abundance of Japanese cars suggests that the photograph is probably one of the cities of the Far East and so on.
the game has a virtual opponent is Alice. She guesses the city, along with the player. We added Alice, that was not boring to play alone. And we wanted to check how well an artificial intelligence to cope with the challenge recognizing cities compared to people. There are probably not obvious for a person details, which one city can be distinguished from the other.
How to prepare photos
the Pictures you see in the game — from Yandex.Cards; downloaded users. We took all of the photos from the Card, made in Russia, and ruled out those that were filmed in the countryside and in small towns. If the city is available only two or three pictures, Alice will not learn to recognize it reliably.
the User photos on Yandex.Maps
Then the pictures are automatically filtered out. First, removed the pictures of poor quality: too small or too dark. Second, the rejected images with inappropriate subjects — for example, photos where most of the frame is a sign or an ad. As a result there are about 400 thousand photos taken in 111 cities of Russia. The distribution of the cities turned out uneven: it is obvious that pictures of St. Petersburg on Yandex.Maps much more than photos of Stary Oskol.
How Alice learned to recognize the city
In the Laboratory of machine intelligence Yandex taught specialized neural network classifier. He didn't just recognize the photo objects, and their features trying to determine what city they could be.
the game involved a neural network with the architecture of the SE-ResNeXt-50 trained on 80 thousand classes. This is the main neural network, which uses Yandex for image processing. For the game, we take the outputs of one of the upper layers of the network and docelem them the classification of images.
the basis of the classifier neural network. Teach her about the same as people: show the picture and tell in which city she removed. All images shown in the training, the neural network learns and finds out in the future with absolute precision.
We have divided the photos into two parts: one was used for training and the other only for games. This allowed us to teach the neural network, and together with her and Alice, to summarize observations and find patterns, not to give the memorized answer.
All the pictures that shows the game, Alice sees for the first time. Moreover, artificial intelligence, unlike people, are not trained in the course of the game. If you play several rounds in a row, some the may fall again — people recognize him, and Alice will look at photos like the first time.
he sees in images neural network
just to understand why the neural network took a decision, it is impossible — but it is possible to recognize areas, which played an important role in the verdict. On a black and white image of the red marked zone that helped the neural network to determine that the photo — Vladivostok (the correct answer).
the Unequivocal answer to the question "Which city on the picture?" the artificial intelligence does not. Instead, it computes probabilities: for example, a neural network may be 75% sure that the picture Yalta and 25% — Moscow.
the Chapel of St. Nicholas in Yalta. Red marked areas that "voted" for that is Moscow. Green — because this is Yalta.
several weeks We have tested the game in Yandex. While statistics show that Alice cope with the guessing on average better than live players. People have a head start — they can determine the city's inscriptions and codes of regions on the car number. Alice in the game does not do this: we deliberately did not add the function of text recognition.
to Start the game
Test yourself — open the game in the browser or say to Alice "let's play "guess the city"". If you manage to bypass the artificial intelligence, you can consider yourself a local expert.
by the Way, you can compete not only with Alice but also with other people — to do this, use the link "challenge friends". The link friends see the same set of pictures as you.
New School managers Yandex 28.03.2019 at 11:20
Yandex + Hyundai Mobis 19.03.2019 at 09:05
Show an example 01.03.2019 at 08:00
Yandex.Mail will eliminate unnecessary mailings 13.02.2019 at 09:00
Yandex invites interns 05.02.2019 at 06:00
Scientific prize named after Ilya Segalovich 31.01.2019 at 07:40
New year in a new format 28.12.2018 at 14:50
Yandex.Remusica 26.12.2018 at 11:15
Speakers-Junior 14.12.2018 at 15:10
Phone with her I 05.12.2018 at 09:05