Abstract:
Object recognition from video is a challenging problem, due to the
inevitable changes in the spatial scale and the temporal length of
object observations across different video sequences. Existing
research fails to address the problem of online recognition, and
semantic representations of video data understandable to a human are
missing. The talk will cover spatio-temporal appearance descriptors of
the first and the second order, a highly configurable family of
descriptors suitable for creating an iteratively refinable
representation of the target object from video data. Also, a semantic
descriptor of video, the COIN descriptor, will be introduced.
Recognition results on a traffic sign dataset, a human action dataset
and a dynamic texture dataset will be presented.
CV:
Karla Brkić received her M. Eng. (dipl. ing.) degree in Computing from
the Faculty of Electrical Engineering and Computing in September 2007.
Since 2008, she has been employed at the same Faculty, as a research
associate at the Department of Electronics, Microelectronics, Computer
and Intelligent Systems, where she works towards a PhD degree. She
participated in a research project funded by the Croatian Science
Foundation, a bilateral Croatian-Austrian research project and a
research project funded by the University of Zagreb Development Fund.
In her research she has been extensively collaborating with the
Vision-based Measurement Group of the Graz University of Technology,
Austria. Since September 2012, she participates in the EU FP7 project
ACROSS. She has co-authored four journal papers and seven papers
published in the proceedings of international conferences. Her
research interests include representation, classification and
reasoning about video data in computer vision.