| NIPS2003 Workshop on
"Open Challenges in Cognitive Vision"
||13 December 2003
Whistler, British Columbia, Canada
- Barbara Caputo, NADA --CVAP/CAS, KTH, Stockholm,
- Henrik I Christensen, NADA --CVAP/CAS, KTH, Stockholm,
- Christian Wallraven, MPIK, Tuebingen, Germany
Basic visual operations such as categorization and complex tasks such
as scene interpretation have long been major challenges for
computational vision. At least some of the these issues call for
integration of methods into systems. Construction of systems for
operation in realistic environments requires integration of methods
from signal processing, geometry, statistics and reasoning.
Naturally a number of both component methods and system
level behaviors can be acquired from studies of biological systems.
Traditionally vision has been studied using a reductionistic
approach. Given the complexity of cognitive tasks, however, it is
not obvious that such an approach is the
most efficient way to address the core problems.
Some issues such as multi-cue figure ground segmentation,
embodied categorization, and behavior / skill acquisition can only be
studied in the context of systems. Recent progress in studies of
categorization, statistical learning theory, active perception,
software engineering and computational neuroscience
is paving the way for improved understanding of
cognitive functionalities in biological and artificial systems. This
workshop will focus on discussion of components methods such as:
The coupling between visual perception, tasks, knowledge and the
visual system requires memory. Issues that are of special importance
for integrating memory into vision systems include:
how to manage representations in presence of limited
resources (such as mechanism for generation of abstractions and forgetting);
model for attention; integration of information across representations and
- Learning and Adaptation:
A system whose goal is that of interacting with
the real world must be capable of learning from experience and adapt to
unexpected changes. Also, there is a need for integration of multiple
visual features to enable generation of stable hypotheses, and for methods
for combination of cues in the presence of uncertainty.
Research has in particular focused on recall of
specific object instances, events and actions. Whereas
recently some progress has been achieved on systems that allow
of object classes, events and scenes across visual appearance,
new methods are needed to enable abstractions and effective
categorization across variations in color, surface markings,
geometry, temporal scenes, context and tasks.
Vision is often considered in isolation. When considered in the
context of an embodied system the concept of an ''active visual observer''
becomes important. The visual system operates here as a task oriented
perception module that generates a diverse set of visual descriptions about
the environment. The set of descriptors is by no means organized in a
hierarchy. Depending on the task at hand the system might generate features
to the ''agent'' in terms of events, labels, and/or spatio-temporal models
(geometry, trajectories, relations, etc). thus, integration plays an
important role from processing of visual cues and multi-modal sensor
fusion to systems architecture.
The goal of the workshop is to present and discuss new methods and techniques
that enables construction of cognitive vision systems that can perform task
oriented categorization and recognition of objects and events.
The focus will be on assesment of state of the art and identification of
Maintained by Barbara Caputo
March 18, 2004