introduction and summary of my experience

With explosion of multimedia applications and widespread access to it, we have abundant data sources in many domains.  The main difficulty is how to interpret and mine useful information for practical solutions.

Machine Learning offer data analysis offer some tools to detect patterns and enable prediction capabilities. I am experienced with

  1. Linear and non-linear (kernel) methods for classification

  2. Multiple information source fusion

  3. Unsupervised learning (clustering, graphs, high-dimensional data)

  4. Semi-supervised learning (semi-supervised SVM, Co-Training)

  5. Probabilistic approaches (graphical models, belief networks)

with application to video and scene understanding, object detection and machine learning.


With aging population in Europe and other developed countries, diseases as Alzheimer and similar are becoming a subject of great concern. The concerns are not purely from the health point of view but also financial that put a burden on social and healthcare services at country level.

The ambitious IMMED project attempts to detect the early signs of the disease and thus reduce significantly the need of time consuming behavior analysis in person. The developed system allows to search efficiently the large corpuses of ego-centric nature HD video recordings for activities and locations in a house. Approved professional studies on the field validated the developed wearable recording prototype and the search software.

My research addressed the issue of image-based localization from large high-quality video recordings. The requirements dictated reliable place recognition in uncontrolled environments at home given weak user annotation.

Every country has a rich cultural and historical legacy in various forms. Handwritten manuscripts are valuable resources that open a whole new world to researchers and ordinary citizens alike. Unfortunately, these documents must be kept in very tightly controlled environments and are thus accessible only by very few scientists. Additionally, their very rich content and amount makes them hard to comprehend and to search.

The DOCEXPLORE project provides three-fold solution to the problem. The valuable documents are not only digitized but also enables to browse the content easily. The developed system provides browsing and annotation tools for the historians and librarians, as well as features rich presentation capabilities for wider public. The final element introduces advanced image-based search capabilities both for researchers and ordinary library or exposition visitors.

My research concerns the design and development of arbitrary selected word and graphical pattern search algorithms. The constraints imply no recognition of words/objects and rapid result delivery to the end user.