ESP-game ( eng. E xtra S ensory P erceptron "extrasensory perception") is a computer game based on the idea of using human computational resources To perform a number of computer science tasks that cannot be solved by a computer ( image recognition for creating complex metadata ), while the recognition task itself is presented in an entertaining way.
The idea of the game and the first version suggested Louis von Ahn , a professor at Carnegie Mellon University . Google Inc. acquired a license to create its own version of the game in 2006, in order to improve the results of the issue in the search for images. [1] The version of the game from Google was closed on September 16, 2011 due to the termination of the Google Lab site in September 2011 .
Content
Idea
Image recognition is a very difficult task for computers. People are quite capable of it, but they will not do it for nothing. Therefore, if we present this task as an entertaining game, people will have a desire to participate. Answering the question about how much the game fascinates the players, the feedback collected from users turned out to be extremely positive.
Possession of a large number of signed images is practically significant; for example, it can be used to more accurately search for images or increase accessibility for visually impaired users (by reading the image name from a computer).
The combination of two people signing the image increases the chances that the words entered will be suitable and accurate. Since the only thing these two people have in common is that they both see the same picture, they have to enter the most appropriate name to coordinate it.
The authors of the ESP game provided evidence that the names derived from the game were really useful descriptions of the pictures. According to the search results of randomly selected keywords, the proportion of relevant images when searching using the names obtained by the game is extremely high. Further evaluation was achieved by comparing the name “generated” by the players with those that were set to describe the image initially.
Game Rules
Having entered the game, each user automatically connects with another random player - opponent. Players do not know each other and can not communicate with each other. After connecting, the players are shown the same image. Their task is to harmonize the word, which would be an appropriate description for the image. They both enter possible names, and at that moment when the same word was entered by both players (not necessarily at the same time), this word becomes the name of the image. When one name becomes consistent, a new picture is shown to the players and the game continues. A total of two and a half minutes is given to name 15 images.
Both players have the opportunity to report a wish to skip the picture (discard). After one of the players "refuses", the second player receives a message that his partner wants to skip the image. Both players must refuse to change the picture to another.
Some images have a taboo title; those. words that cannot be entered as a picture description. These words are usually associated with the image and are designed to complicate the game, because they prevent the use of the simplest “common” words that will be used to designate an image. Taboo words appear during the game itself. If a picture is shown for the first time, then it has no taboo words. If she was once shown, then she has one word-taboo - the name obtained as a result of the first game. The next time you use the picture for the game, it will already have two words-taboo and so on. Taboo words are automatically selected by the system: as soon as an image has been named a sufficient number of times with the same word, it becomes “taboo”; therefore, each image will be described by many different words.
Sometimes a player has to play with a computer, without a human partner. The game itself acts as an opponent, and shows the player a picture with an already defined description (which was collected using previously played games between real people). This is necessary if there are an odd number of people playing at the moment, and it is not possible to divide them into pairs. [2]
At the end of 2008, the game was renamed to GWAP (English “Game W ith A P urpose” - “a game with a goal”), with a new user interface. Some other games that were also created by Louis von Anom, such as Peekaboom and Phetch, were stopped at that moment.
Cheating system
Ahn described countermeasures that prevent players from “cheating” the system by entering deliberately false data. Players are sometimes shown “control” pictures, for which they know in advance what is depicted on them, thanks to which it is possible to check whether the players answered honestly or not. All data entered by the player are taken into account only under the condition that he successfully passed the "test" picture. [3]
In addition, the name is saved only if a certain number of players (N) have indicated it. At this point, the list of word-taboo is cleared, and the picture is sent to the game again, as if it were a completely new image. If X is the probability that the name is wrong, then after N repetitions of games, the probability of hitting the “wrong” name is , provided that the repetitions are independent of each other. [3]
In the current implementation of ESP, the game encourages players to specify “obvious” names for the pictures, which will most likely lead to an agreement with the opponent. But these names can often be obtained from names that are already represented, using the appropriate language model, you just need to add some additional information to the system. Microsoft has developed a research project - a language model that considers the likelihood of the next possibly added name. This model is then used in a program that then plays an ESP game without looking at the image. [four]
Image selection
The choice of the image used in the ESP game affects the player's interest. After all, the game will be boring if all the images are taken from one site and are similar to each other.
When you first run ESP, the game used a collection of 350000 images selected by the developers. Later versions used filtered random images from the Internet. Such images are reintroduced into the game until they are fully signed. [5] Random images were captured using Random Bounce Me, a website that selects a page at random from the Google database. Requests came to "Random Bounce Me" repeatedly, each time collecting all JPEG and GIF images on a random page, except for images that did not meet the criteria: empty, monochrome images, images that are less than 20 pixels in any dimension or with an aspect ratio of more than or less . This process is repeated until 350000 images are collected. Then their size is changed in accordance with the game interface. After that, out of these 350,000 images, a random 15 is selected for each game session.
Links
- ↑ Solving the web's image problem , bbc (May 14, 2008). The appeal date is December 14, 2008.
- ↑ Google Tech Talk on Human Computation by Luis von Ahn , < https://www.youtube.com/watch?v=tx082gDwGcM >
- 2 1 2 Google Tech Talk on human computation by Luis von Ahn
- ↑ Rethinking the ESP Game - Microsoft Research
- ↑ Luis von Ahn. "Human Computation". 2005