Speakers: Michael Wand,
Max-Planck-Institut für Informatik.
Thursday, October 25th, 11:00am, in BC 329
Title: 3D Correspondences and Applications
Abstract:
The talk will discuss recent work that has been done within the
"Statistical Geometry Processing" group at MPI Informatics and Saarland
University. We are working on "shape understanding" algorithms, i.e.,
algorithms that aim at discovering structure in geometric data sets and
utilize it for analysis and modeling. Humans understand shapes already
at an intuitive level. However, finding a formal model that explains
"structure" in shapes to a certain extend is a major scientific
challenge. In addition to being able to capture meaningful aspects, the
models also need to be simple enough to permit efficient and robust
algorithms for discovering such structure in data.
The talk will focus on correspondence analysis as one approach to this
problem: First, we establish correspondences between shapes, i.e.,
detect pieces of geometry that are essentially similar and relate these
to each other. I will discuss various techniques to efficiently and
robustly compute correspondences between shapes, allowing for different
types of invariance. Second, we can go up one level of abstraction and
look at the structure of the obtained correspondences: Assuming we have
discovered multiple, potentially overlapping pairs of regions of
equivalence within a piece of geometry, what does this tell us about the
shape? This question is addressed by "inverse procedural modeling"
techniques that characterize families of shapes that are similar to an
example piece. We use correspondence information to derive shape docking
rules and, alternatively, algebraic invariants to describe such shape
spaces in a constructive manner.