Visual Inference by Data Driven Markov Chain Monte Carlo
Abstract
In this talk, I will present a computing paradigm called DDMCMC for
visual inference. The objective of the paradigm is to provide an
integrated solution to conventional tasks: image segmentation, curve
grouping and recognition in a common framework. The basic idea is to utilize
discriminative model computed in a "bottom-up" process as proposal
probabilities to activate generative models in a "top-down" process and
thus drive the Markov chain for fast convergences. The Markov chain
consists of a number of reversible jumps to search in a complex state
space which contains many subspaces of varying dimensions. I will discuss
various theoretical results on the MCMC hitting-time/mixing-time. One
particular new technique is the Swendsen-Wang Cut (SWC) that we developed
recently. It can generate fair samples on the graph partition spaces with
fast mixing speed. This paradigm is also applied to motion segmentation
and 3D reconstruction from a single image. I will show a number of
experiments in the talk.
This is joint work with three graduate students: Z.W. Tu, F. Han, and
A. Barbu, and a colleague A.L. Yuille.
Bio
Song-Chun Zhu received his Ph.D from Horvard in 1996 (with Dr.
Mumford) and currently he is an associate professor at UCLA, jointly at
Statistics and Computer Science. His research is focused on statistical
modeling and learning of visual patterns and shatistical computing for
visual inference.