Seminar - streda 3.1. od 9:00

[Tomas Bures]

Dobry den,

rad bych vas pozval na zitrejsi seminar:

*Josef Sivic*, Department of Engineering Science, University of Oxford: 
/_Efficient visual search of images and videos_/

/Abstract:/
The objective of this work is to search videos and image collections 
based on their visual content. Despite the recent success of text based 
search (e.g. Google), visual search in unannotated image and video 
collections remains a challenging problem. The imaged appearance of a 
particular object can change significantly due to changing camera 
viewpoint, illumination, and partial occlusion by other objects. 
Furthermore, there is additional variation when searching for visual 
object classes, such as cars, buildings or people, rather than a 
particular object, e.g. the Eiffel Tower.
Inspired by the success of textual search, we have developed a novel 
object representation where an image is described as a collection of 
visual 'nouns', as analogues of words. These descriptors are built with 
a controlled degree of invariance to certain appearance transformations. 
This new representation allows us to employ efficient search methods and 
probabilistic models from statistical text analysis.
In the first part of the talk, I will show rapid retrieval of particular 
objects and people in videos. In the second part of the talk I will show 
how probabilistic topic discovery methods from statistical text analysis 
can be used to learn models of visual object classes from an unlabelled 
collection of images. The learnt models are then applied to object class 
retrieval in image databases and films.
Joint work with Andrew Zisserman, Alexei Efros, Bryan Russell, and 
William Freeman.

-- 
Tomas Bures
Distributed Systems Research Group
Department of Software Engineering
Faculty of Mathematics and Physics
Charles University, Czech Republic
http://nenya.ms.mff.cuni.cz/~bures

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://d3s.mff.cuni.cz/pipermail/seminar/attachments/20070102/043b03e4/attachment.html>