Shiying Li, UNL

Sliced-Wasserstein-based matching and representation of distributional data

Abstract: I will discuss sliced-Wasserstein ideas for matching and representing distributional data. In the first part, I will focus on distribution matching. Although optimal transport (OT) provides a natural framework, its large-scale computation remains challenging, especially in high dimensions. I will describe an efficient iterative slice-matching approach, originally introduced by Pitié et al. for transferring color statistics, in which the resulting sliced optimal transport is built from a sequence of one-dimensional OT problems between sliced distributions. I will discuss convergence properties of such iterative schemes. In the second part, I will show how slicing ideas can also be used to construct feature representations for distributional data to encode certain locally smooth variations. In particular, I will illustrate this through local gradient distributions for face recognition under varying illumination conditions, where lighting-induced deformations can be modeled by a subspace in a suitable transform domain, and recognition is performed by a nearest-subspace method.