Recognition of Shapes by Editing their Shock Graphs
Thomas B. Sebastian, Philip N. Klein, Benjamin
B. Kimia
Brown University, Providence RI, USA
Abstract
We present a novel framework for the recognition of objects based on their
silhouettes. The main idea is to measure the distance between two shapes
as the minimum extent of deformation necessary for one shape to match another.
Since the space of deformations is very high-dimensional, three steps are
taken to make the search practical:
(i) we define an equivalence
class for shapes based on the topological equivalence of their shock graphs;
(ii)
we define an equivalence class for deformation paths based on the
equivalence of their sequence of shock transitions (locus of shock instabilities);
and (iii) we consider deformation paths which consist of a pair
of ``simplifying'' paths from each shape to a common shape. Despite this
discretization of the shape space and associated deformation paths, which
tremendously reduces the search requirement, there still remain numerous
paths to consider. We have developed an edit-distance algorithm for shock
graphs which finds the optimal path in polynomial time, where the cost
of each paths is the sum of the deformation edits to a shock transition.
The resulting correspondences between two shock graphs are intuitive as
illustrated by numerous examples. We examine the robustness of the match
in the presence of an extensive variety of visual transformations
including boundary perturbations, articulation and deformation of parts,
viewpoint variation, illumination changes leading to segmentation
errors, scale changes, and finally partial occlusion, and show that
the matching results are invariant to a significant range of transformations.
The recognition rates on two distinct databases of 99 and 216 shapes each
indicate highly successful within category matches (100% in top three matches),
and render the framework potentially usable in a range of recognition from
shape applications.
Details
Relevant Publications
Acknowledgements
-
The support of NSF Grants CCR-9700146, IRI-9700497, IRI-0083231 and
BCS-9980091 is gratefully acknowledged