PointSet to PointSet registration with random sampling

Roulbac · January 3, 2019, 5:54pm

Hi,
I need to register two point clouds; however my point sets have too many points, so it is quite slow to run.
I have followed the example from the IterativeClosestPoint.
A solution I can think of is to randomly sample my point clouds at each iteration, though I don’t think the registration method allows it…
Any workaround for first randomly sampling my PointSets before registering? I looked around but couldn’t find any. Should I write my own filter? Are there any guidelines/best practices for that?

matt.mccormick · January 8, 2019, 11:32pm

Hi @Roulbac,

Yes, random sampling the points is one good approach.

The points in a PointSet have an integer index. In Python , you could quickly generate integer indices with numpy.random.shuffle and numpy.arange.

In C++, you could reuse the RandomPermutation code used by itk:ImageRandomNonRepeatingConstIteratorWithIndex:

github.com

InsightSoftwareConsortium/ITK/blob/7b0e4702728e70dcb1f7c3c6a91be2c3480e1c55/Modules/Core/Common/include/itkImageRandomNonRepeatingConstIteratorWithIndex.h#L71-L140


class RandomPermutation
{
public:
using GeneratorPointer = typename Statistics::MersenneTwisterRandomVariateGenerator::Pointer;
NodeOfPermutation *m_Permutation;
GeneratorPointer   m_Generator;
SizeValueType      m_Size;


RandomPermutation(SizeValueType sz)
{
  m_Size = sz;
  m_Permutation = new NodeOfPermutation[m_Size];
  m_Generator = Statistics::MersenneTwisterRandomVariateGenerator::New();
  this->Shuffle();
}


RandomPermutation &operator=(const RandomPermutation &it)
  {
    delete[] m_Permutation;
    m_Size = it.m_Size;

This file has been truncated. show original

In either case, create a new point set and populated with points from the complete set using the first N indices.

A bonus challenge is to create a class that encapsulates this operation, and contribute it to the community!

Roulbac · January 9, 2019, 5:49pm

Thank you!
Excited to try and contribute that once I get a bit of a break from school stuff!

keyurs19 · April 10, 2020, 8:16pm

Hello Matt,

Thank you for your post regarding the random sampling on a pointset. Could you please advise how would one uniformly sample the pointsets?

Thank you,
Keyur Shah

dzenanz · April 10, 2020, 9:14pm

I think that two options are full sampling and random sampling. Every n-th point does not make much sense for point sets as it does for images.

keyurs19 · April 10, 2020, 9:46pm

Hello Dzenan,

Thank you for your response. Another quick question, would you know how to write a pointset to a file? I could not find an example to do so.

Thank you,
Keyur Shah

dzenanz · April 10, 2020, 10:44pm

ITK has no specialized classes for reading and writing point sets.

matt.mccormick · April 13, 2020, 7:05pm

For 3D points sets, it is possible to save point sets in mesh file formats like .OBJ. In C++, itk::MeshFileReader / itk::MeshFileWriter can be used. In Python, these are exposed with the more Pythonic itk.meshread / itk.meshwrite functions in ITK 5.1.