To define a dissimilarity measure between geometrical structures, we propose to use a representation of shapes with normal cycles (an object associated with the normal bundle of the shape which encodes all the curvature information). Using kernel metrics on normal cycles, we define a metric between shapes that fits in a framework of inexact registration. It allows for a matching that takes into account the region of high curvature and the boundaries of the shapes.

This presentation is part of Contributed Presentation “CP9 - Contributed session 9”

Authors:

Roussillon Pierre (Ecole Normale Supérieure Paris-Saclay)

Keywords:

image registration, integral equations for image analysis, nonlinear optimization, normal cycles, shape analysis, shape spaces