YOU are lost in a foreign city, you don鈥檛 speak the language and you are late for your meeting. What do you do? Take out your cellphone, photograph the nearest building and press send. For a small fee, photo recognition software on a remote server works out precisely where you are, and sends back directions that will get you to your destination. That, at least, is what two researchers at the University of Cambridge hope their software will one day be used for.
Roberto Cipolla and Duncan Robertson have developed a program that can match a photograph of a building to a database of images. The database contains a three-dimensional representation of the real-life street, so the software can work out where the user is standing to within 1 metre. This is far better than existing systems can manage. GPS satellite positioning is accurate to 10 metres at best, and can be useless in cities where tall buildings shield the user from direct line of sight with the satellites. And positioning using cellphone base stations has a precision of between 50 and 100 metres. 鈥淭elling people 鈥榊ou are in the vicinity of X,鈥� is no good to man nor beast,鈥� says John Craig of Cambridge Positioning Systems, a company that develops software for locating mobile phones.
Unlike the GPS or cellphone base station approaches, Cipolla and Robertson鈥檚 software can tell which direction you are facing. So the service can launch straight into a set of directions such as 鈥渢urn to your left and start walking鈥�, or give information on the building in the photograph.
Advertisement
When their system receives an image it begins by identifying vertical and horizontal lines. Next, it warps the image so that the horizontals are all parallel with each other, and the same for verticals. This transforms the picture into one that was taken square on, rather than at an angle. The software then looks for useful features, such as the corners of windows and doors, and extracts the colours and intensities of the pixels around them. Next, it searches the image database for matching data, using the base station the cellphone鈥檚 signal came from as a guide. Finally, it uses the differences between the two images to calculate the photographer鈥檚 position.
The software can match two images even when they are taken at a different times of day, from different angles and with clutter such as pedestrians and vehicles in the way. 鈥淭hat鈥檚 an easy problem for a human, but it鈥檚 very difficult for a computer,鈥� says Robertson (see 鈥淪nap happy鈥�).
However, the system鈥檚 commercial future is uncertain. 鈥淭he question is: how much are people prepared to pay for it, and how often will they use it?鈥� says Rob Morland, of technology consultants Scientific Generics near Cambridge. 鈥淭hat鈥檚 a tough one.鈥� For now, Cipolla and Robertson are optimistic. Last month they received funding to start working on a prototype to cover all the buildings in Cambridge city centre.
Snap happy
With my digital camera slung round my neck, I could easily be mistaken for another tourist enjoying an English spring afternoon in Cambridge. But while the day trippers will be filling their albums with shots of King鈥檚 College chapel or the Mathematical Bridge, I am after more modest game: the shops surrounding Market Square.
Though my snaps won鈥檛 net me a Pulitzer Prize, they do have a serious purpose. I wanted to put Roberto Cipolla and Duncan Robertson鈥檚 software through its paces and find if it works.
Back in the lab, my photos look very different from those in the database. The buildings are taken from an oblique angle rather than front-on, shop displays are unrecognisable, and I鈥檝e made sure there are plenty of cars and people in the foreground to throw the program off-track.
Unsporting I admit, but real-life users will not care about making things easy; they will just want to know where they are. Despite my efforts, the software was only confused by the most fiendish of my eight snaps, which had a truck blocking around a third of the building.
In Robertson鈥檚 more systematic test, he tried to match 100 photos to a database of the same size. In the three cases where the software fell down, it had either confused two identical buildings or two similar parts of the same building. Updated software could spot this error because the program gives such database images a high similarity score. This could be used as a trigger to prompt the user to take a snap of another building, say, to confirm the fix.
