Geometric Discretization in Shape analysis
Licentiatavhandling, 2022
The first paper considers the problem of finding a parameterized time-dependent vector field that warps an initial set of points to a target set of points.
The parametrization introduces a restriction on the number of available vector fields.
It is shown that this changes the geometric setting of the matching problem and equations of motion in this new setting are derived.
Computational algorithms are provided, together with numerical examples that emphasize the practical importance of regularization.
Further, the modified problem is shown to have connections with residual neural networks, meaning that it is possible to study neural networks in terms of shape analysis.
The second paper concerns a class of spherical partial differential equations, commonly found in mathematical physics, that describe the evolution of a time-dependent vector field.
The flow of the vector field generates a diffeomorphism, for which a discretization method based on quantization theory is derived.
The discretization method is geometric in the sense that it preserves the underlying Lie--Poisson structure of the original equations.
Numerical examples are provided and potential use cases of the discretization method are discussed, ranging from compressible flows to shape matching.
compressible fluids
diffeomorphisms
residual neural networks
quantization
machine learning
Shape analysis
Författare
Erik Jansson
Chalmers, Matematiska vetenskaper, Tillämpad matematik och statistik
Ämneskategorier
Beräkningsmatematik
Annan matematik
Matematisk analys
Utgivare
Chalmers
Pascal
Opponent: Stefan Sommer, University of Copenhagen, Denmark