The MIT researchers tested robotic object recognition in a household setting with everyday objects on a table. Image: Christine Daniloff and Jose-Luis Olivares/MIT

by Scott Kesselman

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have developed a new method for object recognition in household robots that takes advantage of mobility and uses images collected from multiple perspectives to more accurately identify objects.



In an upcoming paper from the International Journal of Robotics Research, the team from MIT shows that by exploiting multiple perspectives, an off-the-shelf algorithm could recognize four times as many objects compared to using a single perspective and also reduced misidentifications.



“If you just took the output of looking at it from one viewpoint, there’s a lot of stuff that might be missing, or it might be the angle of illumination or something blocking the object that causes a systematic error in the detector,” says Lawson Wong, MIT graduate student in electrical engineering and computer science and lead author of the paper. “One way around that is just to move around and go to a different viewpoint.”



Wong and his advisors considered scenarios with 20 to 30 different images of household items grouped closely together on a table with several scenarios including duplicate objects.



The first object recognition algorithm the team tested was developed for tracking systems like radar. The algorithm tracks each object across successive images and determines which objects correspond to which in separate images. This method is time consuming, because every hypothetical correlation identified by the software has to be considered.



To improve the algorithm, researchers sampled images at random and determined only the most likely object in the second group that an object in the first would map to. The algorithm would then reevaluate mappings that did not match in both directions. This method resulted in identifying objects 10 times faster.

<< Back to the News