Shape Spaces
- The shape of a collection of N points is defined as
"whatever is left after the effects of rotation, translation and
scaling have been ruled out" [Kendall, 1984].
If X are the
coordinates of the N points (a 3xN matrix) relative to any reference
frame, and X are the corresponding normalized coordinates (a
point on a 3N-1 dimensional sphere), then a shape s is
obtained by centering and re-orienting the reference
system so as to align it with the principal axes of the cloud of
points.
Therefore, a shape is just a re-positioning of the
coordinates on a special coordinate frame intrinsic to the object
s = p X
p is a rigid change of coordinates (a point on
the Lie group SE(3)) that we call pose .
- A Shape Space can be represented geometrically as the base
of a "fiber bundle", where each fiber describes all (scaled) sets of
points which can be transformed rigidly one to the other.
Kendall, Le, Carne and others have unraveled the Riemannian
geometric structure of shape spaces.
- Problems:
- Object with a different number of points
belong to different shape-spaces
Problems with occlusions.
- Local coordinatization of the shape space
is ill-conditioned in the presence of symmetric
objects [Carne 1992].
4

Stefano Soatto (soatto@vision.caltech.edu), March 1996