Entries with tag real time sketch identification.

Software Identifies Sketches


New software developed by researchers from Brown University and the Technical University of Berlin enables computers to recognize sketches drawn in real time. Computers have been able to match realistic sketches – such as police sketches with mug shots -- but typically have difficulty processing abstract sketches or doodles. The researchers make their point with the example of how people draw rabbits – with exaggerated features such as long ears and a fluffy tail. The images are representational. They don’t look real, but humans know through experience that a particular doodle is meant to represent a rabbit. The key to the technology is a large database of sketches that the researchers used to teach a computer how humans sketch objects. They started by creating a list of everyday objects, then used crowdsourcing to hire people to sketch objects from each category to create the database. The data was used to teach the system based on current recognition and machine learning algorithms. Next, they developed an interface that lets users  input new sketches, which the computer attempts to identify in real time. The system now IDs sketches with roughly 56 percent accuracy; in testing, humans identified the object in the sketch with a rate of roughly 73 percent accuracy. The researchers say the application could help improve sketch-based interface and search applications. They presented their work last month at SIGGRAPH. (PhysOrg)(Brown University)(“How Do Humans Sketch Objects” at Technical University of Berlin website)

Showing 1 result.