Event Location: NSH 1507
Bio: Sinisa Todorovic received the joint B.S./M.S. degree with honors in
electrical engineering from the University of Belgrade, Serbia, in
1994. From 1994 until 2001, he worked as a software engineer in the
communications industry. He earned his M.S. and Ph.D. degrees in
electrical and computer engineering at the University of Florida,
Gainesville, in 2002, and 2005, respectively. Since 2005, he holds
the position of Postdoctoral Research Associate in the Beckman
Institute, University of Illinois at Urbana-Champaign, where he
collaborates with Prof. Narendra Ahuja. Sinisa’s main research
interests concern computer vision, with current focus on unsupervised
extraction and representation of spatial structure in images and
video. He is the recipient of Jack Neubauer Best Paper Award in IEEE
Trans. Vehicular Technology in 2004. He serves as Associate Editor of
Advances in Multimedia.

Abstract: This talk is about solving the following problem: given a set of images
containing frequent occurrences of multiple object categories, learn a
compact, multi-category representation that encodes the models of these
categories and their inter-category relationships, for the purposes of
object recognition and segmentation. The categories are not defined by the
user, and whether and where any instances of the categories appear in a
specific image is not known. This problem is challenging as it involves
the following unanswered questions. What is an object category? To which
extent human supervision is necessary to communicate the nature of object
categories to a computer vision system? What is an efficient, compact
representation of multiple categories, and which inter-category
relationships should it capture?  I will present an approach that
addresses the above stated problem, wherein a category is defined as a set
of 2D objects (i.e., subimages) sharing similar appearance and topological
properties of their constituent regions. The approach derives from and
closely follows this definition by representing each image as a
segmentation tree, whose structure captures recursive embedding of image
regions in a multiscale segmentation, and whose nodes contain the
associated geometric and photometric region properties. Since the presence
of any categories in the image set is reflected in the occurrence of
similar subtrees (i.e., 2D objects) within the image trees, the approach:
(1) matches the image trees to find these similar subtrees; (2) discovers
categories by clustering similar subtrees, and uses the properties of each
cluster to learn the model of the associated category; and (3) captures
sharing of simpler categories among complex ones, i.e.,
category-subcategory relationships. The approach can also be used for
addressing a less-general, subsumed problem, that of unsupervised
extraction of texture elements (i.e., texels) from a given image of 2.1D
texture, because 2.1D texture can be viewed as composed of repetitive
instances of a category (e.g., waterlilies on the water surface).