Joint M-Best-Diverse Solutions for Energy Minimization
Bogdan Savchynskyy
(U. of Dresden, Germany)
Abstract:
We make an overview of recent results in the problem of finding M
diverse low energy labelings for energy minimization in undirected
graphical models. Our approach to infer diverse M solutions jointly
turned out to include as its special cases such standard techniques as
determinant point processes and the sequential DivMBest method of
Batra et al. In spite of the generality and high computational
complexity of the framework in general, we will show several classes
of problems, which can be solved efficiently. These include, but not
limited to diverse solutions for submodular energies. In particular,
we will show a close relation between our diversity method and
parametric submodular minimization/parametric max-flow techniques.