The Push-Forward Transform for Continuous and Robust Comparison of Dynamic Shapes

What the math gives to ML

The paper defines a continuous shape representation by transporting a scalar field on each shape through a diffeomorphism into a shared reference domain. The transferable asset is not merely the use of signed-distance functions, but the explicit separation between intrinsic shape content and nuisance parameterization: once a correspondence map is available, the transported SDF and auxiliary fields can be compared pointwise in canonical coordinates. This suggests a neural shape encoder whose input is a PF-T canonical field rather than raw voxels or arbitrarily aligned point clouds, with an auxiliary deformation module constrained to remain bijective and geometry-preserving.