pclpy: Point Cloud Library for python

Automatically generated python bindings for the Point Cloud Library (PCL) using pybind11 and CppHeaderParser.

The library is in early stage and in active development, do not use in production. The api is likely to change.

Only Windows x64 and python 3.6 are supported at the moment.

Contributions, issues, comments are welcome!

Why another binding?

The python-pcl bindings contain a lot of repeated code. Maintaining, adding features or point types is quite tedious. While Cython is great, binding templated code isn't one of its strenghts. Also, python-pcl implement only a subset of PCL's classes and point types.

Using pybind11, we can use c++ templates directly. The goal is to wrap as much of the library as possible.

Installing

Download PCL release for Windows (PCL-1.8.1-AllInOne-msvc2015-win64.exe) at: https://github.com/PointCloudLibrary/pcl/releases/download/pcl-1.8.1/PCL-1.8.1-AllInOne-msvc2015-win64.exe

pip install pclpy

And tell me if that worked :smile:

Features

• All About half of the point types are implemented (those specified by the default msvc compile flags)
• You can view point cloud data as numpy arrays using cloud.x or cloud.xyz
• boost::shared_ptr is handled by pybind11 so it's completely abstracted at the python level

Example

Here is how you would use the library to process Moving Least Squares. See the PCL documentation: http://pointclouds.org/documentation/tutorials/resampling.php

import pclpy
from pclpy import pcl

point_cloud = pclpy.io.read_las(test_data("street.las"))
mls = pcl.surface.MovingLeastSquaresOMP.PointXYZRGBA_PointNormal()
tree = pcl.search.KdTree.PointXYZRGBA()
mls.setSearchRadius(0.05)
mls.setPolynomialFit(False)
mls.setNumberOfThreads(12)
mls.setInputCloud(point_cloud)
mls.setSearchMethod(tree)
mls.setComputeNormals(True)
output = pcl.PointCloud.PointNormal()
mls.process(output)

You can see it's quite similar to the C++ version:

// C++ version

pcl::PointCloud<pcl::PointXYZ>::Ptr point_cloud (new pcl::PointCloud<pcl::PointXYZ> ());
pcl::io::loadPCDFile ("bun0.pcd", *point_cloud);
pcl::MovingLeastSquaresOMP<pcl::PointXYZ, pcl::PointNormal> mls;
pcl::search::KdTree<pcl::PointXYZ>::Ptr tree (new pcl::search::KdTree<pcl::PointXYZ>);
mls.setSearchRadius (0.05);
mls.setPolynomialFit (false);
mls.setNumberOfThreads (12);
mls.setInputCloud (point_cloud);
mls.setSearchMethod (tree);
mls.setComputeNormals (true);
pcl::PointCloud<pcl::PointNormal> output;
mls.process (output);

Modules

• These modules should work, but are untested
  • 2d
  • common
  • geometry
  • features
  • filters
  • io
  • kdtree
  • keypoints
  • octree
  • recognition
  • sample_consensus
  • search
  • segmentation
  • stereo
  • surface
  • tracking
  • visualization
• These modules are skipped for now
  • ml
  • people
  • outofcore
  • registration
  • every other module not built in the precompiled windows binaries

Not Implemented

(see todo.md and the what to skip section in generators/config.py)

To build

• Download PCL release for Windows (PCL-1.8.1-AllInOne-msvc2015-win64.exe) at: https://github.com/PointCloudLibrary/pcl/releases/download/pcl-1.8.1/PCL-1.8.1-AllInOne-msvc2015-win64.exe
• PCL_ROOT environment variable must be set to the installation directory of PCL
• About requirements:
  • Install pybind11 from github (2.3dev version) it includes a necessary bug fix
  • Install CppHeaderParser from https://github.com/davidcaron/CppHeaderParser (specific bug fixes)
• Generate modules using generate_pybind11_bindings.py
• There is a missing file from the PCL release that you should get from the github repo: 2d/impl/kernel.hpp
• Must be built with x64 version of cl.exe because of the large memory usage (see workaround in setup.py)
• python setup.py install
• Useful setup.py arguments:
  • --msvc-mp-build should enable a multiprocessed build
  • --msvc-no-code-link makes linking much faster (do not use for releases, see setup.py description)
  • --use-clcache to cache msvc builds using clcache (must be installed)
  • --debug to build in debug mode

Roadmap

• Wrap as much of PCL as reasonably possible
• More tests
• CI on Appveyor
• Upload wheels to Pypi
• Make it installable on Linux and Mac