This is a repository with some minimal extensions of the original GenRe Repository to allow for training on different ShapeNet splits. Please refer to the original paper as well. The main focus of this repository is generating ground truth data for training. For setup please follow the instructions in the original repository.

To run our evaluation code, please compile OccNet extension modules in ./eval/mesh_gen_utils

python setup.py build_ext --inplace

Download the training data for the 13/42 split on ShapeNet using the following command. Note that ~620GB of space is required to download this data.

bash download.sh

In the directory where the repository is cloned, make a symlink to where the data was extracted

ln -s path/to/data ./downloads/data/shapenet

Then as described in the original repository, follow the steps to train the GenRe model specifying 13_seen for the class argument.

1. Download data_split.json by running cd util_scripts; wget https://www.dropbox.com/s/7shqu6krvs9x1ib/data_split.json; cd ..

2. Generate symlinks for testing data by running

    python util_scripts/make_genre_test_13_seen_symlink.py
    python util_scripts/make_genre_test_42_unseen_symlink.py
   

3. Run the testing scripts

    bash scripts/test_genre_seen_13.sh
    bash scripts/test_genre_unseen_42.sh
   

4. Postprocess the data to be in synset/object directory structure, since it isn't so by default.

    python util_scripts/postprocess_13_seen_pred.py
    python util_scripts/postprocess_42_unseen_pred.py
   

5. Run the testing scripts in the eval directory

    python eval_seen_13_GenRe.py
    python eval_unseen_42_GenRe.py
   

   and run the following scripts to load the results

    python load_seen_13_GenRe.py
    python load_unseen_42_GenRe.py
   

The pages below contain information to generate ground truth data for GenRe.

1. Image Rendering
2. Full Spherical Images
3. TDF Voxel Grids