TensorFlow Graphics Point Cloud Convolutions

This code developed during the Google Summer of Code 2020 and is pedning in TensorFlow Graphics.

This module contains a python TensorFlow module pylib and a custom ops package in tfg_custom_ops. While it is possible to run without the custom ops package, it is strongly advised to install it for performance and memory efficiency.

Content

This code contains all necessary operations to perform point cloud convolutions

1. Datastructure
   • Point cloud class for batches of arbitrary sized point clouds.
   • Memory efficient regular grid data structure
2. Point cloud operations
   • Neighborhood computation
   • Point density estimation
   • Spatial sampling
     • Poisson Disk sampling
     • Cell average sampling
3. Convolution kernels
   • Kernel Point Convolutions
     • linear interpolation
     • gaussian interpolation
     • deformable points with regularization loss as in KPConv
   • MLP
     • multiple MLPs as in MCConv
     • single MLP as in PointConv
4. Feature aggregation inside receptive fields
   • Monte-Carlo integration with pdf
   • Constant summation
5. Easy to use classes for building models
   • PointCloud class
   • PointHierarchy for sequential downsampling of point clouds
   • layer classes
     • MCConv
     • PointConv
     • KPConv
     • Conv1x1

Installation

Precompiled versions of the custom ops package are provided in custom_ops/pkg_builds/tf_* for the latest TensorFlow versions. For compilation instructions see the README in the custom_ops folder.

To install it run the following command (replace VERSION with your installed TensorFlow version, e.g. 2.3.0)

  pip install custom_ops/tf_VERSION/*.whl

Tutorials

Check out the Colab notebooks for an introduction to the code

• Introduction
• Classification on ModelNet40

Unit tests

Unit tests can be evaluated using

  pip install -r pytest_requirements.txt
  pytest pylib/

These include tests of the custom ops if they are installed.

Additional Information

You may use this software under the Apache 2.0 License.

tensorflow_graphics_point_clouds

Moved to https://github.com/schellmi42/graphics/tree/point_convolutions/tensorflow_graphics/projects

Additional Information

You may use this software under the Apache 2.0 License.