Voxel-Based Convnets for Classification

This is the pytorch implementation of the voxception-resnet as described in the original paper

For the Discriminative model, I've included a MATLAB script in utils to convert raw Modelnet .off files into MATLAB arrays, then a python script to convert the MATLAB arrays into either .npz files or hdf5 files (for use with fuel) or directly as dataset.

The _nr.tar files contain the unaugmented Modelnet10 train and test sets, while the other tar files have 12 copies of each model, rotated evenly about the vertical axis.

the code for training and testinf is present in the file test_and_train.py while the model itself is present in voxception_resnet_pytorch.py

You can produce a simple ensemble by averaging multiple models' predictions on the test sets. I provide six pre-trained models for this purpose, along with .csv files containing their outputs on ModelNet40. Use the test_ensemble.py script to produce a .csv file with the model's predictions, and use the ensemble.m MATLAB script to combine and evaluate all the results.