jackd / shapenet

Python loading/manipulation functions for ShapeNet object dataset.

This repository provides python loading, manipulation and caching functions for interacting with the ShapeNet dataset.

Dependencies:

• Dictionary Interface to Datasets (DIDS)
• util3d

Setup

1. Install pip dependencies

pip install numpy h5py progress wget

2. Clone relevant repositories

cd /path/to/parent_dir
git clone https://github.com/jackd/dids.git
git clone https://github.com/jackd/util3d.git
git clone https://github.com/jackd/shapenet.git

3. Add parent directory to PYTHONPATH

export PYTHONPATH=$PYTHONPATH:/path/to/parent_dir

Consider adding this to your ~/.bashrc file if you do not want to call it for each new terminal. 4. Copy default_config.yaml to config.yaml and make changes depending on where you have data saved etc. See comments in default_config.yaml for more.

cd shapenet
cp default_config.yaml config.yaml
gedit config.yaml  # make changes

Core Dataset

We cannot provide data for the core dataset - see the dataset website for access.

Assign the location of your the ShapeNet dataset to the SHAPENET_CORE_PATH environment variable,

export SHAPENET_CORE_PATH=/path/to/ShapeNetCore.v1

This directory should contain the zip files for each category, e.g. 02691156.zip should contain all obj files for planes.

ICCV 2017 Competition Dataset

To use functions associated with the ICCV2017 competition dataset, after downloading the data,

export SHAPENET17_PATH=/path/to/shapenet2017/dataset

Currently this only supports uncompressed data. This directory should contain train_imgs, test_images, train_voxels, train_imgs, val_imgs and val_voxels directories as provided by the dataset. This may change in the future to support compressed versions.

Data

Most data is provided via a DIDS Dataset. A number of these datasets are saved to disk as required to reduce repeated calculations. These should be calculated and saved as required, but you can manually force the data preprocessing. For example, the following will generate meshes, point clouds and voxels for the plane category in the core dataset with default arguments.

cd core
cd meshes/scripts
python generate_mesh_data.py plane   # parses objs to hdf5 vertices/faces
cd ../../point_clouds/scripts
python create_point_clouds.py plane  # samples faces of meshes
cd ../../voxels/scripts
python create_voxels.py plane        # creates voxels in binvox format
python create_archive.py plane       # zip the binvox files created above

Resulting datasets can be used much like dictionaries, though are required to be explicitly opened or used in a with block.

import numpy as np
from shapenet.core.meshes import get_mesh_dataset
from shapenet.core import cat_desc_to_id
from util3d.mayavi_vis import vis_mesh
from mayavi import mlab

desc = 'plane'

cat_id = cat_desc_to_id(desc)
with get_mesh_dataset(cat_id) as mesh_dataset:
    for example_id in mesh_dataset:
        example_group = mesh_dataset[example_id]
        vertices, faces = (
            np.array(example_group[k]) for k in ('vertices', 'faces'))
        vis_mesh(vertices, faces, color=(0, 0, 1), axis_order='xzy')
        mlab.show()

See examples for more.