README.md

Data

• Download the human segmentation dataset of "Convolutional Neural Networks on Surfaces via Seamless Toric Covers" by Maron et. al. 2017 from here (link by the original authors): https://www.dropbox.com/sh/cnyccu3vtuhq1ii/AADgGIN6rKbvWzv0Sh-Kr417a?dl=0
• Unzip it in to the data subdirectory of this folder like (e.g. on unix run unzip human_benchmark_sig_17.zip -d data/)

Training from scratch

To train the models, use

python human_segmentation_original.py --input_features=xyz  

or, with heat kernel signature features

python human_segmentation_original.py --input_features=hks  

Note that since we do not use a validation set on this dataset (to match prior work), and simply take the accuracy at the last epoch, there is some decent variance in the final accuracy from run to run.

Evaluating pretrained models

Pretrained models are included in /pretrained_models. You can load them and evaluate on the test set like:

python human_segmentation_original.py --input_features=xyz --evaluate  

or, with heat kernel signature features

python human_segmentation_original.py --input_features=hks --evaluate  

results

gt lable

pr label