felixfeliz / bidgcn

BIDGCN

This is the implementation of the boundary informed dynamic graph convolutional network (BIDGCN) proposed in

[1] Carlotta Giannelli, Sofia Imperatore, Angelos Mantzaflaris, Felix Scholz. BIDGCN: Boundary informed dynamic graph convolutional network for adaptive spline fitting of scattered data. 2023. hal-04313629.

The key property of this network architecture for learning on point clouds is a novel input layer that handles boundary conditions defined on a subset of the input point cloud. In [1], the network is applied to the problem of parameterizing scattered 3D point clouds over a planar domain for adaptive fitting with hierarchical splines. This repository also contains the synthetic data set used for training and testing in that publication. Besides point cloud parameterization, the new input layer can potentially be applied to any other problem where boundary conditions are provided.

Notes

The implementation use PyTorch and PyG and can thus be easily integrated into existing codes.

bidgcn_conv.py contains the novel input layer, which is implemented as a subclass of PyG's MessagePassing module.

architecture.py contains the network architecture proposed in [1].

dataset.py contains the code for generating the synthetic training dataset used in [1], it uses the helper functions and classes provided in bezier.py.

radiusedge_conv.py is equal to the PyG-implementation of the dynamic edge convolution operator by Wang et al., 2019 with the difference that a radius graph is used instead of a k-nearest neighbor graph.

training.py contains the training loop for reproducing the network in [1].