Kernel Methods for Geometric Data

Résumé : This talk presents some applications of the KeOps library in utilizing kernel methods for the analysis of large geometric dataset, such as meshes or 3D point clouds (containing up to 10 millions of points) with acquired features. The discussion centers on shape analysis applied to data sourced from medical imaging (where the feature space is a scalar) and spatial transcriptomics (where the feature space is a distribution). A range of generic methods will be introduced for comparing data (possibly across different modalities, as e.g. brain atlases and spatial transcriptomics datasets), through non-rigid deformations and distance definitions in kernel spaces.