Exploiting parallelism in 3D object recognition using the Connection Machine

Abstract

The authors show how data parallelism can be exploited at various stages in the recognition and localization of 3D objects from range data. These stages are edge detection, segmentation, feature extraction; matching, and pose determination. Qualitative classification of surfaces based on the signs of the mean and Gaussian curvature is used to come up with dihedral feature junctions as features for matching and pose determination. Dihedral feature junctions are shown to be fairly robust to occlusion and offer a viewpoint-independent modeling technique for the curved objects under consideration. This offers a considerable saving in terms of storing the object models as compared to the viewpoint-dependent modeling techniques which need to store multiple views of a single object model. Dihedral feature junctions are quite easy to extract and do not require very elaborate segmentation. Experimental results on the Connection Machine showed the advantages of exploiting parallelism in 3D object recognition.<>