Parameterization Invariant Representations for Efficient Shape Learning

Explore a 32-minute conference talk from the Thematic Programme on "Infinite-dimensional Geometry: Theory and Applications" at the Erwin Schrödinger International Institute, where a data-driven framework for parameterization invariant representations of 3D graphs and meshes is presented. Discover how the gradient of the varifold norm enables representations that remain invariant to parameterization while maintaining robustness against sampling noise. Learn about the framework's ability to maintain fixed dimensions regardless of input vertex count, making it compatible with conventional neural network architectures for tasks like classification and registration of raw scan data.