Merrimac uses stream architecture and advanced interconnection networks to give an order of magnitude more performance per unit cost than cluster-based scientific computers built from the same technology. Organizing the computation into streams and exploiting the resulting locality using a register hierarchy enables a stream architecture to reduce the memory bandwidth required by representative applications by an order of magnitude or more. Hence a processing node with a fixed bandwidth (expensive) can support an order of magnitude more arithmetic units (inexpensive). This in turn allows a given level of performance to be achieved with fewer nodes (a 1-PFLOPS machine, for example, with just 8,192 nodes) resulting in greater reliability, and simpler system management. We sketch the design of Merrimac, a streaming scientific computer that can be scaled from a $20K 2 TFLOPS workstation to a $20M 2 PFLOPS supercomputer and present the results of some initial application experiments on this architecture.

This paper presents a new compression scheme for interlaced stereoscopic sequences which differentiates between a region of fixation and a peripheral area, and thereby compacts the stereoscopic information into the spectral space of a monocular video channel. Spectral compression is achieved by avoiding transmitting high-frequency information over the entire images, but only within and around the region where the observer acuity is the highest. The proposed approach consists of decomposing the left and right fields of the stereoscopic pairs into low-pass and high-pass components. High-frequency components are then limited to a fixation region, thus allowing a reduction of their spectral extent. A composite video signal is then formed by positioning the different components into the available spectral space through filtering and modulation. The approach is compatible with the NTSC standard in the sense that the same color subcarrier and the same spectral region are used for the chrominance components. Strategies are proposed for the estimation of the fixation region, based upon a psychophysical study on visual strategies during depth discrimination tasks.

This paper describes a new spectral compaction method for interlaced stereoscopic sequences which combines the stereo information into the spectral space of a single NTSC video channel. The method makes use of the variable spatial and stereoscopic resolution of the human visual system and relies upon the determination of an area of fixation. Spectral compaction is achieved by retaining only the high-frequency information associated with this area. A composite video signal is formed by combining the lowpass, highpass and chrominance components of both channels into the available spectral space by means of modulation and filtering operations. Compatibility of the composite video signal with the NTSC standard is ensured by using the same colour subcarrier for the chrominance components of one channel and a second subcarrier whose phase is inverted on alternate fields to place the chrominance components of the second channel into the Fukinuki holes. Good quality results are obtained with relatively simple separable FIR filters.

This paper describes a new compression method for interlaced stereoscopic image sequences, which compacts the stereo information into the spectral space of a single NTSC channel. The left and right fields of the stereoscopic pairs are each decomposed into a lowpass and a highpass components. High-frequency components are limited to an area of fixation, thus allowing a reduction of their spectral extent. A composite video signal is then formed by positioning the different components in the available spectral space through filtering and modulation. Compatibility of the produced signal with the NTSC standard is ensured by using the same spectral region for the chrominance information and the same colour subcarrier. Good quality results are obtained with relatively simple 2D separable spatial filters at a reasonable computational cost.

The authors consider the problem of detecting global symmetry in images consisting of dense arrangements of local features, such as dots or oriented segments. Images of this kind are encountered in textured scenes or field-type pictorial representations. The proposed computational model, elaborated on the basis of several psychophysical experiments, assumes a precedence of grouping over symmetry detection and consists of three stages: a grouping stage, in which clusters are formed among local elements presenting a sufficient level of mutual affinity; a symmetry detection stage, in which pairs of symmetrical clusters are discovered and their axes of symmetry determined; and a symmetry subsumption stage, in which an attempt is made to detect more global symmetries. An implementation of the model is described. Results are presented showing a good agreement of the model's performance with human symmetry perception.>