Partition - Based Sampling of Warp Maps for Curve Alignment

We propose a flexible sampling method for warp maps used in continuous monotone pairwise alignment of open and closed curves, possibly with landmark constraints. Using the point process machinery, we conduct a detailed study of the sampling method and demonstrate that it prescribes a distribution on the set of warp maps of [0, 1] and the unit length circle S. The distribution (1) possesses the desiderata for decomposition of the alignment problem with landmark constraints into multiple unconstrained ones, and (2) can be centered at a desired warp map. It is based on random partitions of [0, 1] and S and contains a global regularization parameter, both of which enable the sampling of a rich class of warp maps. The distribution can be related to the Dirichlet process on the set of probability measures. Practical utility of the sampling method is demonstrated through (1) a novel stochastic variational algorithm, and (2) a Bayesian model for alignment, for closed and open curves in R, k = 1, 2, 3.