A0160
Title: Robust registration of probability density functions
Authors: Rozenn Dahyot - Trinity College Dublin (Ireland) [presenting]
Abstract: Objective functions originating from optimal transport and information theory frameworks are now widely used in a range of applications, from shape registration, color transfer to machine learning. Registration of functions is also an essential processing step in functional data analysis.This talk is focusing on registering probability density functions by minimizing the robust Euclidean distance L2 or its approximation L2E. We represent probability density functions as Kernel density estimates so that the integral form with L2 has an explicit expression, and the resulting objective function is optimized efficiently with standard simulated annealing algorithms. It is shown to be robust and flexible allowing to take into account correspondences when available. For illustration, it is applied to shape registration and color transfer for image processing and computer vision applications.
