ETL-0456 ADA 18 3 755 hJ 1

We present solutions to a set of benchmark problems for Image Understanding for the Connection Machine. These problems were proposed by Darpa to evaluate architectures for Image Understanding systems, and are intended to comprise a representative sample of fundamental procedures to be utilized in Image Understanding. The descriptions of solutions on the Connection Machine embody several general methods of using the machine to filter images, to determine connectivity among image elements, to determine geometry of image elements and, finally, to compute graph properties, such as matchings and shortest paths. Mike Drumheller, Willie Lirm, Guy Blelloch, Carl Feynman, all of Thinking Machines Corporation, and Todd Cass, of the Al Lab, have all been instrumental in contributing algorithms and good ideas about using the Connection Machine. Vision System The parallel computing environment at the MIT Al Lab consists of a Connection Machine [4] with 16K processsors with a Symbolics 3640 Lisp Machine as host. The Connection Machine The Connection Machine (CM) [41 is a parallel computing machine having 64K processors, operating under a single instruction stream broadcast to all processors (figure 1). Each of the processors is a simple 1-bit processor with 4K bits of memory. There are two modes of communication among the processors: first, the processors are connected by a mesh of wires into a 256 x 256 grid network (the NEWS network, so-called because of the four cardinal directions), allowing rapid direct communication between neighboring processors, and, second, the router, which allows messages to be sent from any processor to any other processor in the machine. The processors in the Connection Machine can be envisioned as being the vertices of a 16-dimensional hypercube. Figure 2 shows a 4-dimensional hypercube; each processor is connected by 4 wires to other processors. Each processor in the CM is identified by a unique integer in the range 0... 65536, its hypercube address, imposing a linear order on the processors. This address denotes the destination of messages handled by the router. Messages pass along the edges of the hypercube from source processors to destination processors. To allow the machine to handle data with more than 64K elements, the Connection Machine supports the concept of virtual processors where a single physical processor can operate as multiple virtual processors by serializing operations in time. The number of virtual processors assigned to a physical processor is denoted ;~~ I, """':':;",;"".ii." . ./? .i:. :" " """" Figure 1: Block Diagram of the Connection Machine (from [4]) Figure 2: 4-dimensional Hypercube

[1]  Leonidas J. Guibas,et al.  Parallel computational geometry , 1988, 26th Annual Symposium on Foundations of Computer Science (sfcs 1985).

[2]  Michael Ian Shamos,et al.  Computational geometry: an introduction , 1985 .

[3]  W. Daniel Hillis,et al.  The connection machine , 1985 .

[4]  Ellen C. Hildreth,et al.  Comments on "Digital Step Edges from Zero Crossings of Second Directional Derivatives" , 1985, IEEE Trans. Pattern Anal. Mach. Intell..

[5]  Robert M. Haralick,et al.  Digital Step Edges from Zero Crossing of Second Directional Derivatives , 1984, IEEE Transactions on Pattern Analysis and Machine Intelligence.

[6]  Jan van Leeuwen,et al.  Maintenance of Configurations in the Plane , 1981, J. Comput. Syst. Sci..