For those who need to train the model on their own dataset, here is our voxelization process.

1. Download and compile this mesh voxelization library, and put the executable file voxelize under this directory.

2. Organize the PartNet mesh data as follows, e.g.,

   Bed/
      45/ # shape id
         00.obj # part mesh, possibly not watertight
         01.obj
         ...
         04.obj
      ...
      10967/
         00.obj
         01.obj
         ...
         16.obj
      ...

   It's important to name each obj file as xx.obj(xx being the index of part) as it is required when running voxelization.

3. Convert .obj to .off and scale the shapes from [-1, 1]^3 to [0, 64]^3.

   python obj2off_scaled.py -i Bed -o Bed_off

4. Run voxelization. run_voxelize.sh. This will voxelize each shape folder (containing part .obj files) into a single h5 file (containting voxels of dimension (n_parts, 64, 64, 64)).

   bash run_voxelize.sh Bed_off Bed_off_h5 0 64

5. Run post-processing. This will post-process the above h5 files, e.g. fill inside voxels and scale each individual part to a $64^3$ canonical space.

   bash run_postprocess.sh Bed_off_h5 Bed_off_h5_post

   After the program finishes, Bed_off_h5_post will contain h5 data that can be used for training.