In So Kweon received the B.S. and the M.S. degrees in Mechanical Design and Production Engineering from Seoul National University, Seoul, Korea, in 1981 and 1983, respectively, and the M.S. and Ph.D. degree in Robotics from the Robotics Institute at Carnegie Mellon University, Pittsburgh, U.S.A, in 1986 and 1990,respectively. He worked for Toshiba R&D Center, Japan, and joined the Department of Automation and Design Engineering at KAIST in 1992. He is now a Professor in the Department of Electrical Engineering at KAIST.
His research interests are in computer vision, robotics, pattern recognition, and automation. Specific research topics include invariant based visions for recognition and assembly, 3D sensors and range data analysis, color modeling and analysis, robust edge detection, and moving object segmentation and tracking. He is a member of ICASE, IEEE, and ACM.

Some selected relevant papers include:
"Robust Model-based Scene Interpretation by Multilayered Context Information", Computer Vision and Image Understanding (CVIU), 2006.
"Object Recognition using Generalized Robust Invariant Feature and Gestalt Law of Proximity and Similarity", 5th IEEE Workshop on Perceptual Organization in Computer Vision (in CVPR'06), New York NY, 2006.
"Biologically Motivated Perceptual Feature: Generalized Robust Invariant Feature", LNCS 3852:305-314 (ACCV'06), 2006.
"Scene Interpretation: Unified Modeling of Visual Context by Particle-based Belief Propagation in Hierarchical Graphical Model", LNCS 3852:963-972 (ACCV'06), 2006.
"3D Target Recognition using Cooperative Feature Map Binding under Markov Chain Monte Carlo", Pattern Recognition Letters, 27 (7): 811-821, 2006.

In this tutorial, we present our experiences in applying machine learning techniques to some notable vision problems, such as object categorization, super resolution, and low-level image processing.
Human vision system can categorize ten thousands of object classes easily. How can computers or robots also have this intelligence Recently, many researchers have tried to explore this problem. However, it is difficult due to the large intra class variation. At the same time, the visual context provides relational information among visual components such as part, object and scene. In general, modeling both the relation and uncertainty (variation) is very difficult problem. Fortunately, graphical model can to this job by combining both areas of graph theory (relation) and probability theory (uncertainty), and provide powerful, flexible framework for representing and manipulating global probability. In this presentation, we introduce the techniques of graph model based approach for categorization.
We briefly introduce the basic concepts of graphical model theory by examples, including directed graphical model (Bayesian Network) and undirected graphical model (Markov Random Field), etc. Then we move to the vision problem and show that Graphical model is suitable for modeling the contexts in vision, followed by some examples of graphical model in categorization in three main parts.
The first part introduces the constellation model for object categorization, which is a well designed model to describe the relation of different parts of object categories in terms of appearance, relative location and relative scale.
In the second part, we mainly introduce the examples of generative Markov Random Field and discriminative Conditional Random Field for modeling hierarchical context in scene. Some other related main scene analysis examples are also briefly introduced.
The final part introduces some graphical models which originate from text modeling techniques and their intrinsic relations, including Bag of Words for object categorization, pLSA and LDA for scene categorization. They model the hierarchical relations between local features, objects and scene.
We also try to introduce some of our research experience about how the BP(Belief Propagation) functions as a solution of the smoothness constraints in example-based super-resolution. Super-resolution is a way of obtaining the super-resolved image from the low resolution image(s) and example-based super-resolution achieve it through two steps, training phase and synthesis phase. In training phase, low resolution and corresponding high resolution training patch pairs are stored. We account for super-resolution problem briefly and demonstrate the effect of BP with a few experimental results.