This is the official implementation of the paper Scalable and Efficient Functional Map Computations on Dense Meshes.
This paper relies on my pyFM package, which you can install or simply clone (it is very lightweight).
The demo also uses my VisualizationTools repo but that's simply for visualization, and is not needed in the main codebase large_mesh.py.
To use the repo, clone using
git clone --recurse-submodules [email protected]:RobinMagnet/Scalable_FM.git
You can run Demo.ipynb to see code for spectrum approximation and shape matching.
The only nonstandard dependency is the pymeshlab package, only for its implementation of Poisson Disk Sampling.
The code follows the notations in the paper, and all functions have docstrings for documentation.
- It requires to work around some surprising behaviours of scipy.sparse.csgraph.dijkstra when using maximal distance after which to stop. This function indeed always returns a dense matrix, with NaN at points further away than the given radius instead of a sparse one. Since this dense matrix usually doesn't fit in memory, we first subdivide the shape in several parts using K-Means and BallTree in 3D, and compute a local dijkstra on each subshape
- Upgrading the radius requires some tricks in order not to recompute all entries at each step.
If you have any question about the code, don't hesitate to reach out to me at robin.magnet [at] inria.fr
If you use our work, please cite us
@article{magnetScalableEfficientFunctional2023,
author = {Magnet, Robin and Ovsjanikov, Maks},
year = {2023},
title = {Scalable and Efficient Functional Map Computations on Dense Meshes},
journal = {Computer Graphics Forum},
publisher = {The Eurographics Association and John Wiley & Sons Ltd.},
}