Robots would have an easier time identifying objects if they moved around while taking stock of their environment, according to researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL).
In a new study, the researchers explain that a standard algorithm for aggregating perspectives recognized four times as many objects as a single-perspective algorithm. Additionally, using multiple perspectives reduc
ed the number of 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, a graduate student in electrical engineering and computer science and lead author on the new paper. “One way around that is just to move around and go to a different viewpoint.”
The study also describes a new algorithm developed at CSAIL, that is as accurate as the multiple-perspective algorithm but works up to ten times as fast. In order to develop robots that can work in the home, along side humans, speed is an important consideration. Household robots need to be able to function fast enough to keep up with their owners.
For the study, the researchers considered a scenario involving 20 to 30 household objects packed close together. Some objects appeared multiple times and images were gathered from many perspectives. Each algorithm must try to identify corresponding objects in multiple images and generates hypothesis about which object from one image might be the same object seen in another image. The challenge lies in quickly assessing all of the many hypothesis generated when many images are used.
The first algorithm tested was a well-respected algorithm used for tracking systems such as radar. This algorithm attempts to speed things up by discarding all but its top hypotheses at each step.
The new algorithm developed at CSAIL kept all the hypothesis it generated, but randomly canvassed them instead of assessing them all. The researchers also kept the number of hypothesis they generated low by considering each object in the first image separately and evaluating the likelihood of it corresponding to an object in the second image, instead of considering every possible set of matches between the two images.
More information on the algorithms will appear in the forthcoming issue of the International Journal of Robotics Research.