The extraction of information from live or off-line video media is becoming evident. In most cases the definition of “what we would like to detect” or “what we can detect” is difficult. On the other hand the content analysis is increasingly indispensable because of the huge number of installed cameras.

The Short-term Visual Memory suits to satisfy the needs of adaptive real-time content indexing and retrieval tasks. This smart indexing and searching system is able to insert video frames (marked with ROI mask optionally) into the database and immediately recall the frames with similar visual content. The effectiveness comes from the limited number of indexed frames. This limitation depends on the hardware resources: the algorithm stores as many information as can be handled optimally. The support of region-of-interest (ROI) allows of the event driven usage: system integrators can define objects or image parts to be indexed.

Thanks for the advanced database implementation the searching queries are accomplished under near constant time duration. The resultant data set is available in a sorted list ranked by the matching of the frame content.


Application areas

  • Video surveillance
  • Object tracking
  • Object searching
  • Visual associative memory
  • Media indexing


  • Preliminary object seg-mentation
  • Dedicated computational power


  • Input from video recording or live camera
  • Real-Time performance
  • Build indices on-line
  • Effective memory usage
  • Robust object descriptions
  • Multi-scale processing
  • Event triggered searching
  • Suitable for multi-camera systems
  • Available in both C and C++ library

Bolck diagram of iSVM visual short term memory
iSVM - Visual short term memory system


download: pdf


Contact: MTA-Sztaki, Kende u. 13-17, 1111 Budapest, Hungary
Tel: (+36) 1 279 6106