Event Location: NSH 1507
Bio: Minwoo Park received the B.Eng. degree in electrical engineering from Korea University,Seoul, in 2004, and the M.Sc. degree in electrical engineering from The Pennsylvania State University in 2007. He is a Ph.D. candidate in the Department of Computer Science and Engineering at The Pennsylvania State University. The common thread of his research has been in understanding the theory and application of a probabilistic graphical model on computer vision problem. His particular interests are in automatic understanding of 3D from an image, perceptual grouping, event recognition, and an efficient inference algorithm.

Abstract: Probabilistic graphical models (PGMs) are widely used in many areas of computer
vision and machine learning. Since the classic belief propagation is not suitable
for a large or a continuous state space, sampling-based belief propagation methods
have been developed, e.g. non-parametric belief propagation (NBP). However NBP requires a large number of samples and its resampling process is too slow, preventing its wide applicability. Therefore, we are motivated to develop an efficient belief propagation method called mean-shift belief propagation (MSBP). MSBP is a heuristic method that works iteratively with local weighted samples to
infer max-marginals within a large or continuous state space.

Furthermore we propose a novel data-driven MSBP (DDMSBP) based on recent work on smoothing-based optimization and mean-shift, that is more robust at finding a significant mode of the marginal density than the original MSBP. In addition, we show that inferencing is in linear time with respect to the number of samples used in MSBP for arbitrary unary potential functions as long as the pair-wise compatibility functions remain Gaussians.

We demonstrate the effectiveness of our novel MSBP and DDMSBP method through simulations and by applying them to challenging computer vision problems. Given a perfect conceptual match between PGMs and the underlying structures of repeating patterns, we first apply our proposed inference algorithm to the detection of near-regular patterns in real-world images. Second, we utilize detected lattice for urban scene analysis. Lastly, we apply DDMSBP to 3D neuroimage de-formable registration as a 3D graph inference scheme.