Registration (for curves, surfaces or images) can be seen as a representation of a shape dataset as a subset of the diffeomorphism group, each shape in the dataset being associated with an optimal deformation of a template. We will present some of these methods, focusing on those that compute this deformation by optimizing a distance over diffeomorphisms under the constraint of mapping a template to a target, and such that the minimized distance is in turn a metric in shape space. These “metric registration” methods include a large spectrum of algorithms deriving from the “large deformation diffeomorphic metric mapping” (LDDMM) framework, and “metamorphosis” in which non-diffeomorphic shape variations are also allowed. We will describe mathematical foundations and algorithms, and the various forms that these methods may take in relation to applications.
Back to Geometry and Learning from Data Tutorials