A Review of 3D Point Clouds

Parameterization Methods
Zaiping Zhu, Andres Iglesias, Lihua You and  Jian Jun Zhang 
2022/06/23
Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
What is 3D point clouds parameterization?
• It is the process mapping a 3D point cloud onto a suitable (usually simpler)
domain.

Parameterization

Original 3D point clouds Point clouds in the parameterization domain

Why point clouds parameterization matters?
Many applications

• Parametric surface reconstruction

• Texture mapping
• Object classification
• …
Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
Point clouds types

• Organized (structured) point clouds

The point cloud data are stored in a structured manner.

Similar to a 2-D matrix and its data are divided into rows and columns according
to the spatial relationships between the points.

• Unorganized (unstructured) point clouds (mostly)

Stored arbitrarily.

A collection of 3-D coordinates, each of which denotes a single point.

Parameterization types

• Global parameterization
Whole point clouds is mapped.

No seams in the reconstructed surfaces, but computationally expensive.

• Local parameterization
Point clouds is partitioned into multiple subsets, each of which is mapped.

Computationally cheap, but seams may exist

Topological shapes

Based on the number of holes they own

genus-0: no holes Genus-1: one hole

genus-2: two holes Genus-3: three holes

https://en.wikipedia.org/wiki/Genus_g_surface#Genus_0
Mapping ( based on what to be preserved)

Suppose f is a bijective function between a mesh S or a point cloud and a

mapping domain S∗, then f is:

• Isometric mapping: length preserving

• Conformal mapping: angle preserving

• Equiareal mapping: area preserving

Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
To parameterize organized point clouds

A topologically identical 2D triangulation from the underlying 3D

triangulation of the point cloud is obtained, which determines the parameter
values of the vertices in the domain plane. Depending on the ways of
transforming from 3D to 2D:

• Harmonic parameterization

• Floater’s barycentric mappings

• Isometric parameterization
To parameterize organized point clouds

Energy function has also been defined to minimize the metric distortion in the
transformation from 3D to 2D: They normally map the boundary points first,
then the inner points are mapped by minimizing defined energy function.

• Edge based energy function

• Spring energy function

Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
Categories
Based on the property of the mapping process:

• Base surfaces methods

• Meshless parameterization method

• Spherical mapping

• Methods adapted from mesh parameterization

• Neural networks based

• Others
Base surfaces methods
Input point clouds

Base surfaces: approximate the underlying

structure of point clouds
• Can be a plane, Coons patch, or a cylinder...
• The parameter values of each point can be
Base surface
obtained by projecting the point cloud onto Base-surface based Local
parameterization
a base surface.
• The projection can either be perpendicular to the surface or based on a
determined projection vector.
Properties of base surfaces

• Unique local mapping: two different original 3D points should be

mapped onto two different locations on the mapping domain.

• Smoothness and closeness: be as smooth and simple as possible,

while still approximating the underlying surface as much as possible.

• Parameterization of base surface: We can choose a more suitable

way to parameterize a base surface by referring to the underlying
structure of the fitting surface.
Ways to obtain base surfaces

• “Simplicial” surfaces: obtained using contouring algorithm after

defining the sign distance to the point clouds
Their method is capable of reconstructing a surface with or without boundary
from an unorganized point set
There is no formal guarantee that the reconstructed result is correct and the
space required to store the reconstruction is relatively large
Ways to obtain base surfaces
• From boundary curves: automatic or manually, some characteristic
curves approximating the underlying structure of the point cloud are
sufficient in defining a base surface
Just four corner points can be used to create a base surface in some cases
Inner points information is not considered
Ways to obtain base surfaces
• Dynamical base surfaces: gradually improved regarding its
approximation to the underlying structure of the point clouds
DBS
Recursive DBS
DBS + recursive subdivision
Base surfaces comparison
Meshless Parameterization methods
Mapping
In order to make sure the
reconstructed mesh has high
quality, the mapping points should
preserve the local structure of the
(a) Input point clouds (b) Meshless parameterization
original point clouds as much as
possible. Therefore, the shape
Triangulation
distortion ought to be minimized in
the parameterization process. This
is formulated as the problem of
Same triangulation
edge structure solving a sparse linear system.

(c) Delaunay triangulation of

(d) surface triangulation the mapping points
Meshless Parameterization methods
 Spherical mapping
(a) A bulldog point clouds
(b) The spherical parameterization
of the bulldog point cloud
The reason why “spherical mapping” is
applied when the underlying structure of the
point clouds is closed (there are no
boundaries) can be partly explained by the
uniform theorem, which states that every
genus-0 closed surface is conformally
equivalent to . Thus, mapping from a genus-
0 surface to the unit sphere is natural. The
same idea is also applied to genus-0 point
clouds.

: the spherical mapping points

: the original point set
: is the centre of the original point set
: represents the largest distance between the original
point set and the centre with the radius of the sphere.
Spherical mapping-application: meshing
Similar to meshless parameterization, Spherical parameterization is mostly used to mesh point
clouds. Directly triangulating a point cloud is challenging, especially for complex geometry,
which can be achieved more easily with the aid of spherical conformal parameterization.

Specifically, instead of directly triangulating a point cloud, the points on the unit sphere after
mapping are triangulated using the spherical Delaunay triangulation algorithm. Then triangulation
of the original point cloud can be obtained from the triangulation on the spherical point cloud as
these two point clouds have a one-to-one correspondence.
Spherical mapping comparison
Adapted from mesh parameterization
• Tewari et al. modify the harmonic one-form method used in parameterizing manifold
meshes to parameterize genus-1 point clouds that are sampled from such meshes. They
locally parameterize the subsets of a point cloud and the way they parameterize the point
cloud can guarantee the consistency between the pieces.

• Li et al. present a new method to reconstruct quad-dominant mesh from unorganized point
clouds using the adapted periodic global parameterization method, which is modified
from the periodic global parameterization method that is used to parameterize a triangle
mesh.
Adapted from mesh parameterization
Neural networks based
Adaptive sequential learning Radial basis
function(RBF) networks

It has a simple topology linking an input layer, a

hidden layer and an output. The input layer has
an i-dimensional input , the hidden layer has K
kernels , k = 1, 2, ...K, and the network output is
a linear combination of kernels.

Fast optimization that minimizes

parameterization residuals

Feed-forward RBF network topology.

 Neural networks based
Residual deep neural networks
Since the network approximates the function
that assigns a suitable parameter value to a
sequence of data points, optimal curve
reconstruction from point clouds can be obtained.

However, their method is only applicable to a

small number of sample points. And the public
data set is not available, so their method is not
very general to real data.

(a) The layout of a building block (b) the layout of the whole residual
neural network
Neural networks comparison
Others
• Many works consider mapping them onto a simple domain with a fixed boundary shape such as a sphere, a
circle or a rectangle. However, some undesirable distortion may occur during the parameterization process
due to the fixed boundary shape. To overcome such a problem, Choi et al. develop a free-boundary
conformal parameterization technique to parameterize disk-shape point clouds, which leads to high quality
of the reconstructed mesh. By free boundary, it means that the positions of only two boundary points are
fixed, and the left boundary points are parameterized to a suitable location automatically based on the
structure of the original point clouds.

• Zwicker et al. present a system in which interactively parameterizing point clouds can be done. During the
mapping process, an objective function is applied to minimize distortions automatically. Furthermore, the
user can adjust the mapping intuitively at the same time.
Others
Outline

1. Introduction to 3D point clouds parameterization

2. Some concepts

3. Parameterization methods of organized point clouds

4. Parameterization methods of unorganized point clouds

5. Conclusion
Conclusion
• It should be pointed out that there is no “best” parameterization method applicable
to all point clouds, as one method may succeed in parameterizing some point
clouds but fail in parameterizing other point clouds. Therefore, for a given point
cloud, it is necessary to choose a suitable method to parameterize the 3D point
cloud according to the desirable properties of low distortion and high computing
efficiency in parameterizing the point cloud.
• Parameterizing point clouds using neural networks technique is a
potential direction, such as data set creation, curve and surface
reconstruction from point clouds using Neural network.
Thank you for your
attention !