Anyone who has typed an address into Google Street View has probably had the experience of seeing that the result shows the street, but not always the exact building at a specified address. This is because computer-vision software hasn’t been advanced enough to zero in on numbers. This is namely due to the fact addresses are displayed on buildings in a variety of colors, sizes and fonts, making it difficult for computers to pinpoint, recognize and extract information from them.
However, researchers at Stanford University have teamed up with Google to improve the technology by creating an algorithm that’s able to more accurately identify street numbers. In “Reading Digits in Natural Images with Unsupervised Feature Learning,” the teams explain how they trained the system to recognize these numbers using computer-vision algorithms combined with technology that recognizes patterns and learns to adapt to and implement them.
It is the hope of Google that this kind of information could lead to a better Street View system, in addition to more accurate maps and navigation services.