Deep Learning Modules

Adaptive Geodesic Convolution (AGC)

The Adaptive Geodesic Convolution (AGC) module provides a way to perform convolution operations on meshes, by computing geodesic patches around each vertex and applying a learnable filter to these patches. The size of the patches is also learned during training, allowing the model to adapt to the local geometry of the mesh.

Biharmonic Distance

This module calculates the biharmonic distance between points on a 3D mesh. Unlike standard geodesic pathways, biharmonic distance provides a smooth, geometry-aware distance metric that captures the global shape of the mesh.

Note

Reference: This module is a direct implementation of the methodology detailed in Biharmonic Distance by Yaron Lipman, Raif M. Rustamov, and Thomas A. Funkhouser (ACM Transactions on Graphics, 2010).