FMNet.pytorch

A pytorch implementation of Deep Functional Maps (FMNet).

Introduction

This is a pytorch implementation of Deep Functional Maps. Groundtruth labels of FAUST correspondence are not used. For efficiency, 2048 points are randomly sampled from 6890 points on original meshes. The results may not be bijective.

Update: Visualization pairs of KeyPointNet are post-processed by PMF to be bijective, while faust pairs are not.

Usage

Build shot calculator:

cd utils/shot
cmake .
make

Calculate eigenvectors, geodesic maps, shot descriptors of trained models, save in .mat format:

python preprocess.py

Train:

python train.py --dataset=faust

Test(temporarily use trained data to test, for visualization):

python test.py --dataset=faust --model_name=epoch300.pth

Visualize correspondence:

python visualize.py

Visualization