Neural Adjoint Maps for refining shape correspondences

What is it about?

In this work, we tackle shape matching by moving beyond the classical assumption of linear representations. We introduce a neural version of functional maps that doesn’t rely on linearly alignable embeddings. Instead, we overfit small neural networks to learn flexible maps, robust even in challenging cases where assumptions like isometry or template similarity break down.

Featured Image

Why is it important?

Accurate shape matching is a cornerstone in geometry processing, with applications from 3D reconstruction to medical imaging. Classical methods rely on assumptions like near-isometry or linear alignment, which often fail in real-world data. Our method lifts these constraints, enabling more robust and generalizable correspondences — a step toward more flexible and powerful geometric understanding

Perspectives