Parallel Scalable Libraries and Algorithms for Computer Vision

Research Assistants: Y. Liu, J. Xu

Sponsor: DARPA

This project develops libraries, algorithms, and tools to enable the effective use of scalable high performance computing technologies. The research focuses on the application area of computer vision and image processing (CVIP). The research is predicated on the philosophy that high performance computing must meet seemingly contradictory goals: software development should be carried out in an architecture- and technology-independent environment, but both algorithms and system software should take full advantage of the hardware features of potentially diverse architectures. These contradictory demands are reconciled by pursuing research in three areas:

• The development and validation of a computational model that serves as a platform for architecture-independent algorithm development, gives an accurate prediction of algorithm performance on a particular architecture, and supports a realistic scalability metric.

• The implementation of fundamental algorithms and complete CVIP task scenarios on diverse architectures, including the MasPar MP-1 and MP-2, Intel Touchstone Delta, iPSC/860, Paragon, and the IBM SP-1 and SP-2.

• The development of applications-driven library-based software tools that bridge the software gap between usability and high performance. Emphasis is on using libraries and a computational model to capture scalability information, and on providing applications users with high performance codes in a flexible environment.