VEIL: Research in Knowledge Representation for Computer Vision — Final Report
Information Sciences Institute, University of Southern California
Abstract
The VEIL (Vision Environment Integrating Loom) project focused on integrating
advanced knowledge representation (KR) technology with image understanding
technology. VEIL to developed a more declarative approach to the construction of
vision systems and produced a tool that incorporates that methodology. Systems
were constructed in a more principled fashion that made it possible to share and
reuse software across systems. Experiments in two main areas were carried out.
We first demonstrated the utility of using Loom as a software engineering tool
for a specific vision application (runway detection). We also demonstrated the
benefits Loom provides for image understanding itself (event detetion).
The major innovations in this work are as follows:
- applied a methodology that maximizes use of declarative knowledge (as
opposed to procedural knowledge) in vision systems, thereby enabling us to apply
modern software development techniques. The criteria for recognizing objects
was stated explicitly in a formal language (instead of being buried in code) making
it easier to understand and maintain an application and keep it consistent.
Extending the recognition capabilities of the software was made easier.
- use of this declarative system construction methodology to facilitate the
process of integrating high-level vision routines (such as for recognizing
sequences of scenes) with low-level routines that recognize picture elements.
- enabling interaction with the system at a level of abstraction appropriate
to the domain task. This includes associating collateral information with the
objects recognized by low-level image understanding programs.
- development of a foundation for a vision ontology.
This work leveraged off the Loom Knowledge Representation system. Loom
captures the best features of object-oriented programming, data-driven
programming, problem solving, and constraint programming, through the use of
an underlying logic-based representation scheme. This system is a powerful tool
that incorporates very strong, frame-based representation capabilities, explicit
term subsumption, and a number of powerful reasoning paradigms (including
logical deduction, object-oriented methods, and production rules). Loom also
provides knowledge representation integrity through consistency checking, and
provides truth maintenance. Infusing these facilities into the vision problem
area, where strong KR capabilities have not yet been developed will significantly
alter and improve the methodology for the construction of vision systems. We
also developed spatial and temporal reasoning capacities (critical for vision), along
with mechanisms to exercise flexible control strategies and incremental scene
processing. Finally, Loom was interfaced to a variety of vision processing
elements to provide a new tool of extended capabilities. The net result is a
powerful software environment for the development of vision systems.
The full paper is available in PDF. (68pp, 2.9MB)
Back to Paper List