On the Local form and transitions of the shock graph for object recognition
Speaker: Benjamin B. Kimia , Division of Engineering, Brown University
Date: April 21 2004
Time: 2:00PM to 3:00PM
Location: 32-D507
Host: Greg Shakhnarovich, CSAIL
Contact: Greg Shakhnarovich, gregory@ai.mti.edu
Relevant URL: ABSTRACT:
The use of a suitable shape representation is critical for a number of
visual tasks. We describe how the shock graph, a dynamic hierarchical
representation of the medial axis, is used for object recognition from
silhouettes. Our approach is based on capturing the topology of the shape space
via a dis-similarity metric that is the cost of the optimal deformation
path between two shapes. Since the space of deformation paths is
infinite-dimensional we discretize it by defining equivalence classes
based on the shock graph topology and its transitions, which are
related to the classical instabilities of the medical axis. A formal
analysis of the local form of the shock graph and its transitions under
a one-parameter family of deformations is described. The
transition-based description of deformation paths is then searched
under an edit-distance paradigm to find the optimal path. We describe
recognition results which are stable under a range of visual
transformations and which for several databases of up to 1032 shapes
recover the correct category.
See other events that are part of MIT Machine Vision Colloquium Spring 2004
See other events happening in April 2004
