Current sketching software only recognises the geometry of a sketch - for example, that two lines meet at a specific angle - and cannot recognise context at all. The MIT software monitors the image as it is being drawn onto a computer screen and allocates probabilities to various interpretations of what it might represent. As the user adds more detail, the software adjusts these weightings. To do this, it uses a technique known as Bayesian analysis, which is normally used to compute the likelihood of specific causes, given certain effects.
"With our software, the 'causes' are what the user had in mind to draw, and the 'effects' are what was actually drawn," says Randall Davis, who has developed the code with Christine Alvarado at MIT's Artificial Intelligence Laboratory. Once the computer has settled on an interpretation, it applies the laws of physics such as gravity and friction to set the virtual world in motion. Cars roll down hills, pendulums swing, and objects bang into each other. The Bayesian part of the software is still at an early stage: it recognises only squares. Davis's next goal is to get it to recognise a much broader range of crudely drawn shapes. Part of the problem is thinking up all the possible ways someone might draw a shape, so as to be able to assign realistic probabilities to particular patterns on the screen.
Written by Celeste Biever
UK CONTACT - Claire Bowles, New Scientist Press Office, London:
Tel: 44-207-331-2751 or email firstname.lastname@example.org
US CONTACT - Michelle Soucy, New Scientist Boston Office:
Tel: 617-558-4939 or email email@example.com
New Scientist issue: 13th September 2003
PLEASE MENTION NEW SCIENTIST AS THE SOURCE OF THIS STORY AND, IF PUBLISHING ONLINE, PLEASE CARRY A HYPERLINK TO: http://www.
"These articles are posted on this site to give advance access to other authorised media who may wish to quote extracts as part of fair dealing with this copyrighted material. Full attribution is required, and if publishing online a link to www.newscientist.com is also required. Advance permission is required before any and every reproduction of each article in full - please contact firstname.lastname@example.org. Please note that all material is copyright of Reed Business Information Limited and we reserve the right to take such action as we consider appropriate to protect such copyright."