VAMPIRE___Visual Active Memory Processes and Interactive REtrieval
VAMPIRE Events Publications Consortium Media archive
Intro Research Activities Scenario 'Mobile augmented reality' Scenario 'video annotations' Slideshow
object recognition and learningvisual trackingaction recognitionAR gearself localisationscene analysis
contextual analysisinteraction and augmented realitysystem integration

Embodiment by "Human in the Loop" paradigm

Embodiment is one of the key concepts in cognitive vision research. It describes the ability of the system to acquire knowledge of its world from the consistent coupling of its own action and perception. Thereby, observing a grasping action can be related to performing the same kind of action and, thereby, is becoming meaningful for the system. This kind of perception-action-cycle is thought of as a fundamental basis for cognitive development.

The VAMPIRE project aims at developing technology for interactive cognitive assistance. In contrast to autonomous systems, robotic actions are compensated by a tight human-system interaction putting the user in the processing loop of a perception-production-cycle using augmented reality techniques. The system shares the same view with the user. Perceptions trigger augmentations of the common view. The user reacts on these by changing the view or explicit feedback to the system. In turn the system uses the feedback of the user in order to adapt its internal representations.

Thus, in the "Human-in-the-Loop" paradigm, cognitive development is driven by

  • the interacting user
  • the dynamic environment
  • an adaptive cognitive architecture

Visual Active Memory (VAM) Concept

The visual active memory concept unifies three main ideas in a single approach

  1. coupling of model acquisition and recognition processes
  2. component-based architectures
  3. self-organizing systems
The coupling of model acquisition and recognition processes leads to a hybrid approach integrating data-centered (persistent memory) and process-centered (consolidation and re-encoding) principles. The memory is structured into different levels of abstraction
  • Pictorial data is used for the acquisition of object models
  • Recognition results are stored in an episodic memory that provides data for the acquisition of contextual models
  • In turn the contextual models can be used to validate memory content, to generate higher-level interpretation, or to trigger user interactions or learning processes
A major challenge in cognitive vision research is the question how to construct vision systems that provide a large spectrum of functional capabilities (e.g. localization, tracking, classification, visualization, prediction, categorization, learning ...). The VAM provides the infrastructure for component-based architectures:
  • Different components are decoupled by mediating the information flow between processes through the active memory. A high degree of transparency allows distributed processing.
  • A unifying XML data model for data storage and communication
  • Trigger and event mechanisms for coordinating different kinds of processes (bottom-up as well as top-down)
The self-organization enables the system to actively control its own resources and to adapt to new tasks. This is realized by
  • a self-descriptive XML data model
  • an internal process environment (including e.g. generic forgetting and compacting)
  • treating module configurations as memory elements
  • a hypotheses concept that includes meta-data, like reliability values and time-stamps

Agile Project Management

  • SnipSnap
  • Frequent integration
  • Close collaboration with partners
  • People exchanges