(via Mathias Eitz)
Computer programs have been written to dominate humans in games such as chess and Jeopardy. Now a program has been created which is on the path to give humans a challenge in Pictionary. Researchers from the Berlin Institute of Technology and Brown University in Providence have used a database of drawings to help a computer correctly identify and categorize human sketches. The program will be the first of its kind to allow computers to have a “semantic understanding” of abstract sketches.
Computers have already been used for accurate search and match applications such as matching mugshots with police sketches. However, a sketch of something such as a cartoon bunny, which children could easily decipher, is too complicated for computers to identify. To overcome this challenge, the researchers put together a database consisting of 250 object categories. Using Mechanical Turk, a crowdsourcing market place through Amazon, 20,000 sketches where collected to make up the categories. Integrating the sketches with existing machine learning algorithms and recognition software, the team created an interface to identify humans sketching in real time.
The program cannot dominate humans in Pictionary just yet, identifying only 56 percent of images, while humans managed about 73 percent. However, expanding the database and possibly collecting data through a free phone app, the program may eventually reign supreme. As of now some of the guesses the computer will output can be amusing. For example, it recognized teddy bears as helicopters and skeletons as pineapples. Other than Pictionary games, the researchers stated the program has strong potential for future sketch-based interfaces and search applications.
A version of this program is available for iOS Called "WhatsMySketch."
Cabe