Mesh classification

We performed our experiments on meshes using the Manifold40 dataset. Please contact us to get access to our prepared dataset. In all our experiments we use hesiod to handle configs. Every time that you run a command, you will be prompted with a UI asking to set paths and other needed configs.

Create the dataset of INRs

The first step consists in creating the dataset of INRs. To do that you can run:

$ cd mesh_classification 
$ python3 create_inrs_dataset.py

(Optional) Compute SDF for an existing dataset of INRs

The first step (Create the dataset of INRs) saves the precomputed SDF for each shape inside each INR h5 file. If for any reason you have a dataset of INRs without the precomputed SDF, you can create a new version of the dataset containing also the precomputed SDF by running:

$ python3 preprocess_sdf.py

Train inr2vec

After creating a dataset of INRs, you can train inr2vec on it by running:

$ python3 train_inr2vec.py

If you need to stop the training and restart it later, you can run:

$ python3 train_inr2vec.py /PATH/TO/INR2VEC/RUN.YAML

where /PATH/TO/INR2VEC/RUN.YAML provides the path to the config file created by hesiod for the training of inr2vec.

Visualization

Aftet training, you can visualize meshes reconstructed by inr2vec running:

$ python3 viz.py /PATH/TO/INR2VEC/RUN.YAML

where /PATH/TO/INR2VEC/RUN.YAML provides the path to the config file created by hesiod for the training of inr2vec.

Export embeddings

After training inr2vec, its encoder can be used to embed the INRs into compact embeddings, by running:

$ python3 export_embeddings.py

Classification

To perform the classification experiment, simply run:

$ python3 train_classifier.py