Keynote speakers

We are proud to announce the following keynote speakers at SCIA 2013.

Maja Pantic

Professor, Imperial College London, Computing Dept., UK and University of Twente, EEMCS, Netherlands.

Abstract

Facial behaviour is our preeminent means to communicating affective and social signals. There is evidence now that patterns of facial behaviour can also be used to identify people. This talk discusses a number of components of human facial behavior, how they can be automatically sensed and analysed by computer, what is the past research in the field conducted by the iBUG group at Imperial College London, and how far we are from enabling computers to understand human facial behavior.

Ethem Alpaydin

Professor, Department of Computer Engineering,┬áBogazi├ži University, Istanbul.

Abstract

As in other branches of science and engineering, in machine learning too, we carry out experiments to get information about the process under scrutiny. In our case, this is a learner, which, having been trained on a dataset, generates an output for a given input. An experiment is a series of tests where we play with the factors that affect the output. These factors may be the algorithm used, the training set, input features, and so on, and we observe the changes in the response to be able to extract information. The aim may be to identify the most important factors, screen the unimportant ones, or find the configuration of the factors that optimizes the response—for example, classification accuracy on a given test set.

Our aim is to plan and conduct machine learning experiments and analyze the data resulting from the experiments, to be able to eliminate the effect of chance and obtain conclusions which are statistically significant. In machine learning, we target a learner having the highest generalization accuracy and is robust, that is, minimally affected by external sources of variability.

In this talk, first, we are going to discuss strategies of experimentation, basic principles of experiment design, and statistical procedures and tests which we can employ to analyze the resulting experimental data. In the second part, we will see some recent work we did on statistical comparison, namely, the MultiTest method where (time/space) complexity is used together with generalization accuracy in comparison, and new tests for comparing kernel machines using hinge/epsilon-sensitive loss.

Related Publications
– E. Alpaydin (2010) Introduction to Machine Learning, 2nd ed, MIT Press (Chapter 19).
– O. Irsoy, O. T. Yildiz, E. Alpaydin (2012) “Design and Analysis of Classifier Learning Experiments in Bioinformatics: Survey and Case Studies,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, 9(6), 1663-1675.
– O. T. Yildiz, E. Alpaydin (2012) “Statistical Tests using Hinge/Epsilon-Sensitive Loss,” ISCIS 27, Paris, France.
– A. Ulas, O. T. Yildiz, E. Alpaydin (2012) “Cost-Conscious Comparison of Supervised Learning Algorithms over Multiple Data Sets,” Pattern Recognition, 45(4), 1772-1781.

Anders Heyden

Professor, Department of Mathematics (LTH), Lund Institute of Technology / Lund University.

Abstract

Robot vision has been an active field of research since the pioneering book by Horn. Due to the latest development within robot technology, implying smaller and cheaper robots and within camera technology, it is still of highest relevance. The main problem related to vision is to use cameras to map-making, navigation, collision avoidance and interaction with the environment. In order to do so we need to know the relative orientation and position of the camera with respect to the robot, the so called hand-eye calibration.

In this talk I will review the standard methods for hand-eye calibration and present some novel methods based on multilinear constraints and solving polynomial equations. I will also describe the map-making problem and illustrate with novel research based on covariance propagation. Finally, I will illustrate some applications within navigation and positioning.

Jiri Matas

Professor, The Center for Machine Perception, Department of Cybernetics, Faculty of Electrical Engineering, Czech Technical University, Prague.

Abstract

The problem of finding correspondences in images arises in many computer vision problems. Many factors contribute to the level of complexity of the problem, the class of geometric and photometric constraints that hold between the images being the most prominent one, together with the level of occlusion.

In the talk, we will mainly focus on the evolution of the methods that aim at finding correspondence under large (extreme) changes of viewpoint. We will discuss the standard pipeline and show that various recent improvements in almost all stages significantly increase the range of solvable problems without large computational costs.

Finally, we will present some matching problems that are beyond the capabilities of current matching methods.