Georgia Institute of Technology have built a database of sketches and photographs that learns how to match basic drawings with real objects.
We previously explored the potential of a search engine that works on image recognition rather than text, and now computational researchers are taking that further by matching simple hand sketches to digital photographs of the object.
The team from Georgia Institute of Technology have developed the Sketchy Database, which consists of 75,000 simple hand-drawn sketches and 12,500 photos of various animals and objects. Using a convolutional network and machine-learning, the team developed software capable of matching, for example, a badly drawn elephant to photos of elephants with 37 percent accuracy. While still in development, the software has much potential for businesses — law enforcement could pair witness sketches to criminal databases, and platforms could enable foreigners, or those who can’t read or write to use sketching as a way of communicating or accessing online content.
How could sketch recognition software benefit other digital networks?