README.md

Creating the ground-truth 3D point clouds

• ShapeNet

  The surface point clouds for ShapeNet are readily available from the AtlasNet repo. Run this script to download and extract the point clouds of all object categories (19G). You should have the ShapeNet data (from the main README) downloaded.
  Under this directory (data), run

  python3 shapenet_create_pointcloud.py

  A separate pointcloud3.npz file will be created for each CAD model in the corresponding folder in NMR_Dataset. You would need to modify the script, mostly in the first few lines:

  • atlasnet_path: where ShapeNetV1PointCloud was extracted.
  • categories: all ShapeNet category IDs to process.
  • num_proc: number of parallel processes.

• PASCAL3D+

  We use the virtual scanner to sample point clouds from only the outer surface of the CAD models.
  You can either follow the installation instructions from the official repo, or alternatively, you can follow the below steps to install via conda.

  First, create a new conda environment via

  conda env create --file vscan-requirements.yaml python=3

  This creates a conda environment named vscan-env. Activate it with

  conda activate vscan-env

  Under this directory (data), clone the repo containing the virtual scanner package and compile:

  git clone https://github.com/wang-ps/O-CNN
  cd O-CNN/virtual_scanner
  python3 setup.py install

  If all goes well, you should be able to import ocnn in Python. Check for errors by running python3 -c "import ocnn".

  Next, run make to compile a simple parser that converts the virtual scanner data format to PLY files. An executable named parse will be created.

  Finally, run

  python3 pascal3d_create_pointcloud.py

  The corresponding virtual scanner output *.points, the converted PLY (*.ply) and Numpy (*.npy) files will be created for each CAD model in the corresponding folder. You would need to modify the script, mostly in the first few lines:

  • root_original: where the PASCAL3D+ CAD models (in OFF format) are stored.
  • root_processed: where the processed point clouds will be stored.
  • categories: all PASCAL3D+ category names to process.
  • num_random_points: number of random points to densely sample.