MVILAIN@G.BBN.COM (Marc Vilain) (02/25/88)
BBN Science Development Program AI Seminar Series Lecture OBJECT RECOGNITION USING MASSIVELY PARALLEL HYPOTHESIS TESTING Lewis W. Tucker Thinking Machines Corporation Cambridge, MA (TUCKER@THINK.COM) BBN Labs 10 Moulton Street 2nd floor large conference room 10:30 am, Tuesday March 1 Problems in computer vision span several layers of data representation and computational requirements. While it is easy to see how advances in parallel machine architectures enhance our capability in "low-level" image analysis to process the large quantities of data in typical images, it is less obvious how parallelism can be exploited in the "higher" levels of vision such as object recognition. Traditional approaches to object recognition have relied on constraint-based tree search techniques that are not necessarily appropriate for parallel processing. This talk will introduce a model-based object recognition system designed at Thinking Machines Corporation and its implementation on the Connection Machine. The goal of this system is to be able to recognize a large number of partially occluded objects in 2-D scenes of moderate complexity. In contrast to previous approaches, the system described here utilizes a massively parallel hypothesize-and-test paradigm that avoids serial search. Perceptual grouping of features forms the basis for generating hypotheses; parameter space clustering accumulates weak evidence; template matching provides verification; and conflict resolution ensures the consistency of scene interpretation. Results from experiments with databases ranging from 10 to 100 objects show the relative independence of the time to recognize objects with either the complexity of the scene or number of objects in the database. -------