8 Parallel Implementation of Algorithms Havaldar Worked on Perceptual Grouping for Object Rec- Ognition. A. Francois Worked on Case-based Reasoning for Recognition. D. Kim Worked on Indoor Navigation. N. Milhaud Worked on Navigation from Infrared Images

knowledge representation technology can be applied to computer vision and to explore extensions to knowledge representation systems to directly aid computer vision research (especially in the area of spatial reasoning). The VEIL project investigates the benefits available to vision applications obtainable via the introduction of declarative programming techniques, specifically, techniques available using advanced symbolic processing technology found in a modern knowledge representation system. In typical vision applications today, a programmer invents specialized data structures and carefully crafts a suite of vision processing algorithms that exploit those data structures. The result is most often a highly specialized piece of code that cannot be reused for a different domain, or applied to applications other than the one originally intended. The Image Understanding Environment project [Mundy et al. 1993] addresses the issue of basic data structures and processing but does not deal with higher level representation issues that are addressed here. This project aims to develop a technology whereby much of the work that goes into the development of specialized vision processing modules results in software that can be shared or reused by multiple applications. Knowledge representation techniques have been a part of computer vision from the beginning [Winston 1975, McKeown et al. 1985, Draper et al 1989]. The difference in this project includes the power of the knowledge representation system being used and the use of relatively mature computer vision programs and techniques as the basis for incorporating knowledge representation technology. This work is reported more fully in [Price et al. 1994]. We have recently initiated a project on parallel implementation of some of our IU algorithms on a Thinking Machines CM-5 computer. Our first task is parallel implementation of the monocular building detection system described in Section 6. The first step of this system, as well as many of our other IU systems, is linear feature extraction. Even though linear feature extraction is considered a low level task, we know of no parallel system implementation for this task. This has been the focus of our initial effort. Linear feature extraction consists of several stages: edge detection, edge linking and line approximation. Edge detection is the process that has been typically studied for parallel implementations. The other steps, those of linking and approximation, however pose some challenges as the data structures change from being " iconic " to symbolic (lists). We have successfully implemented a version of the USC line finder [Nevatia …