D.
Lombardi: Models and methods for unbalanced optimal
transport
Abstract: The unbalanced optimal transport problem is
considered, which is of interest when applying optimal
transport principles to real data analysis. Indeed, real
data are not, often, densities whose mass is conserved. In
order to apply an optimal transport principle, the mass
conservation constraint has to be relaxed by adding mass
sources. Among the several possibilities, eulerian models
of sources are investigated. A general functional form of
the source is derived by requiring that the fundamental
properties of optimal transport are satisfied and, when
the mass is conserved, the problem collapse to the
Benamou-Brenier formulation. As a result, the
interpolation is not featured by an infinite speed of
propagation. The numerical discretisation proposed is
based on an explicit adaptation of the Benamou-Brenier
method. Several test-cases are described in order to
validate the approach and assess its numerical properties.
Moreover, some examples on interpolation of medical images
are proposed.
|