Exploiting parallelism in 3-D object recognition using the connection machine

Bhandarkar, S.M., Shankar, R.V. and Suk, M., Exploiting parallelism in 3-D object recognition using the Connection Machine, Robotics and Autonomous Systems, 8 (1991) 291-309. This paper addresses the problem of parallelism in 3-D object recognition and localization using a fine-grained SIMD architecture such as the Connection Machine TM. The input images are range images of partially occluding three-dimensional objects. The objects are made up of piecewise compositions of curved surfaces. The surfaces of interest are cylindrical, conical and spherical surfaces since they cover a large portion of objects encountered in an industrial environment. Qualitative classification of surfaces based on the signs of the Mean and Gaussian curvature is used to come up with dihedral feature junctions. Dihedral feature junctions are shown to be robust to occlusion and offer a viewpoint independent scheme for modeling the object models and are well suited for matching and pose determination. Hough clustering is chosen as the constraint propagation mechanism on account of the ease of parallelization. Issues regarding parallelism on the Connection Machine TM for every stage of the recognition and localization process such as segmentation, feature extraction matching and pose determination are discussed. Experimental results on the Connection Machine TM bring out the advantages of exploiting parallelism for 3-D object recognition and localization on a fine-grained SIMD architecture.