Shape reconstruction from point clouds using closed form

solution of a fourth-order partial differential equation

Zaiping Zhu1, Ehtzaz Chaudhry1, Shuangbu Wang1, Yu Xia1, Andres Iglesias2,

Lihua You1 and Jian Jun Zhang1
1The National Center for Computer Animation, Bournemouth University, UK
2 Department of Applied Mathematics and Computational Sciences, University of Cantabria,
Cantabria 39005, Spain

Abstract. Partial differential equation (PDE) based geometric modelling has a

number of advantages such as fewer design variables, avoidance of stitching ad-
jacent patches together to achieve required continuities, and physics-based na-
ture. Although a lot of papers have investigated PDE-based shape creation, shape
manipulation, surface blending and volume blending as well as surface recon-
struction using implicit PDE surfaces, there is little work of investigating PDE-
based shape reconstruction using explicit PDE surfaces, specially satisfying the
constraints on four boundaries of a PDE surface patch. In this paper, we propose
a new method of using an accurate closed form solution to a fourth-order partial
differential equation to reconstruct 3D surfaces from point clouds. It includes se-
lecting a fourth-order partial differential equation, obtaining the closed form so-
lutions of the equation, investigating the errors of using one of the obtained closed
form solutions to reconstruct PDE surfaces from differential number of 3D
points.

Keywords: Shape Reconstruction, Fourth-order Partial Differential Equation,

Closed-form Solutions, Error Analysis.

1 Introduction

Shape reconstruction has a lot of applications in many fields. Various surface recon-
struction methods have been developed. These methods include polygon-based, im-
plicit surface-based, and parametric surface-based. In addition, soft computing is also
introduced into parametric surfaces to improve shape reconstruction from point clouds.
Shape reconstruction uses polygon meshes, implicit surfaces, and existing paramet-
ric surfaces such as Bézier, B-spline, and NURBS surfaces has some weaknesses. They
include bid data, heavy geometry processing, high data storage cost, and slow data
transmission over computer networks. How to address these weaknesses is an unsolved
topic.
In contrast, PDE-based shape reconstruction has the following advantages. First, a
single PDE surface patch can describe a complicated shape leading to smaller data than
NURBS, polygon and subdivision modelling techniques. Second, adjacent PDE surface
patches naturally maintain position, tangent, or higher continuities and no manual
2

operations are required to stitch different PDE patches together. Third, any irregular
boundaries can be quickly specified by drawing a closed curve on 3D models, different
sculpting forces can be applied to achieve the expected shape, and global shape manip-
ulations can be easily obtained through shape control parameters etc., leading to more
efficient shape manipulations.
However, a main difficulty for PDE-based shape manipulation is how to solve partial
differential equations. Due to this difficulty, most studies investigated implicit PDE-
based shape reconstruction which involves numerically solving partial differential
equations. Although some research studies investigated explicit PDE-based shape re-
construction by interpolating four curves or satisfying the constraints on two opposite
boundaries of a PDE patch, few studies presented closed form solutions of partial dif-
ferential equations for 4-sided PDE patches. In this paper, we will propose a mathemat-
ical model, derive its closed form solutions, and use one of the closed form solutions to
achieve shape reconstruction from point clouds.

2 Related Work

There are a lot of work of investigating shape reconstruction from point clouds. A com-
prehensive literature survey has been made in [1, 2]. Among these shape reconstruction
methods, PDE-based shape reconstruction has also been investigated. The existing
shape reconstruction methods can be divided into polygon-based, implicit surface-
based, explicit surface-based, and soft computing and parametric surface-based. Shape
reconstruction using explicit surfaces from solutions to partial differential equations
was summarized in [3]. In this section, we briefly review some work on shape recon-
struction from point clouds.
Polygon-based shape reconstruction is most popular. Many of them are based on the
Delaunay triangulation. For each initial border edge in triangulated reconstruction,
Boissonnant estimated a tangential plane and took a vertex of a surface triangle to be
the sample point which maximizes the angle between the vertex and the 𝑘-nearest
neighbors projected to the tangential plane [4], Hoppe et al. used 𝑘-nearest neighbors
to find a tangential plan of every sample point and the marching cubes algorithm and
the signed distance of the sample point closest to the tangential plan to reconstruct 3D
surfaces [5]. Oblonšek and Guid presented a procedure to triangulate the input scattered
point set, extract features, and fair the triangular mesh to achieve surface reconstruction
[6]. Bernardini et al. proposed a ball-pivoting algorithm to interpolate a given point
cloud to reconstruct a triangle mesh where a triangle is formed with three points when
a ball touches them [7]. Gopi et al. projected the neighborhood of each of sample points
on a tangential plan, derived the 2D Delaunay triangulation on the tangential plan, and
mapped the result back to the 3D space, and reconstructed a 3D surface from the map-
ping [8]. Lee et al. reconstructed a 3D surface through repeated subdivision, displace-
ment, and resampling of a control mesh model [9]. Jeong and Kim constructed a coarse
base mesh from a bounding cube containing the input point cloud, and successively
subdivide, smooth, and project it to the input point cloud to obtain shape reconstruction
[10]. Not fitting dense smooth polygonal surfaces, Nan and Wonka used simple polyg-
onal surfaces for reconstruction of piecewise planar objects [11].
 3

Implicit surface-based shape reconstruction has also been extensively investigated.

Duan et al. proposed a PDE-based deformable surface to reconstruct 3D models from
volumetric data, unorganized point clouds and multi-view 2D images [12]. Linz et al.
developed new formulations and fast algorithms to conduct implicit partial differential
equation-based surface reconstruction [13]. Franchini et al. used the numerical tech-
nique based on an efficient semi-implicit scheme and finite volume space discretization
to solve a time-dependent partial differential equation for implicit shape reconstruction
[14]. Pana and Skala introduced an energy functional to combine flux-based data-fit
measures and proposed a regulation term and a continuous global optimization method
to carry out surface reconstruction from an oriented point cloud [15]. Using an implicit
and continuous representation of reconstructed surfaces and optimizing a regularized
fitting energy, Liu et al. developed a level-set based surface reconstruction method to
process point clouds captured by a surface photogrammetry system [16]. He et al. pre-
sented two fast algorithms, one uses the semi-Implicit method and the other is based on
the augmented Lagrangian method, to improve computational efficiency of surface re-
construction from point clouds [17].
Some researchers investigated explicit surface-based shape reconstruction. In order
to solve the heavy computational cost of finite element methods or finite difference
methods in solving the elliptic partial differential equation for 3D shape reconstruction,
Li ang Hero developed a fast spectral method to improve the performance [18]. Ugail
and Kirmani proposed an explicit partial differential equation-based reconstruction
method which interpolates four curves parametrized in terms of the parametric variable
𝑣 to obtain a reconstructed surface [19]. This method was also used to obtain recon-
struction of 3D human facial images [20]. Rodrigues et al. obtained an analytical solu-
tion of a Laplace equation and used it in 3D data compression and reconstruction [21].
Sheng et al. integrated a point cloud update algorithm, a rapid iterative closest point
algorithm, and an improved Poisson surface reconstruction algorithm together to im-
prove the efficiency of surface reconstruction [22].
Soft computing has also been introduced into parametric surfaces to optimize shape
reconstruction. Iglesias et al. introduced function networks into B-spline surfaces to
solve the problem of shape reconstruction [23]. Gálvez and Iglesias integrated an iter-
ative two-step genetic-algorithm and polynomial B-spline surfaces for efficient surface
reconstruction from point clouds [24]. They proposed a particle swarm optimization
approach to reconstruct non-uniform rational B-spline surfaces from 3D point clouds
[25].
Since few research studies investigated shape reconstruction using explicit PDE sur-
faces, we will tackle this issue in this paper. The PDE mathematical model and its
closed form solutions will be investigated in Section 3. The proposed PDE-based shape
reconstruction and error analysis are examined in Section 4. Finally, the conclusion is
drawn and some future research directions are discussed in section 5.

3 Mathematical model and closed form solution

Partial differential equation-based shape reconstruction can be roughly divided into two
categories: one uses implicit solutions of partial differential equations and the other
4

uses explicit solutions of partial differential equations. Shape reconstruction using im-
plicit solutions of partial differential equations involves a lot of numerical calculations,
causing slow shape reconstruction which is not suitable for many applications requiring
real-time performance. In contrast, shape reconstruction using explicit solutions of par-
tial differential equation is based on accurate analytical or approximate analytical solu-
tions of partial differential equations, which involves fewer calculations and is more
efficient than shape construction using implicit solutions of partial differential equa-
tions.
However, a main problem for shape reconstruction using explicit solutions of partial
differential equations is how to obtain accurate analytical or approximate analytical
solutions of partial differential equations. Since solving partial differential equations
analytically is not an easy task, the current explicit solutions of partial differential equa-
tions used for PDE-based geometric modelling and shape reconstruction mainly deal
with two boundaries of a PDE surface patch, i. e., accurately satisfy partial differential
equations and continuity constraints at two opposite boundaries of a PDE surface patch.
How to obtain accurate analytical solutions of partial differentia equations which ex-
actly satisfy partial differential equations and continuity constraints on four boundaries
of a PDE surface patch is an important topic.
A vector-valued partial differential equation used to describe a 3D surface patch in-
volves two parametric variables 𝑢 and 𝑣. The four boundaries of the 3D surface patch
are defined by 𝑢 = 0, 𝑢 = 1, 𝑣 = 0, and 𝑣 = 1. In order to satisfy positional continui-
ties, four unknowns should be included in a closed form solution to a vector-valued
partial differential equation to satisfy four positional functions, i. e. boundary curves at
the four boundaries of a 3D surface patch. Similarly, in order to satisfy up to tangential
continuities, eight unknowns should be involved in a closed form solution of a vector-
valued partial differential equation to satisfy four positional functions and four tangen-
tial functions at the four boundaries of a 3D surface patch. From the theory of partial
differential equation, the closed form solution to a second-order partial differential
equation of parametric variables 𝑢 and 𝑣 has four unknowns, and the closed form solu-
tion to a fourth-order partial differential equation has eight unknowns.
Up to tangential continuities is most popularly used to create smooth 3D models.
Taking all of these factors and a closed form solution into account, we propose to use
the following vector-valued fourth-order partial differential equation for shape recon-
struction
𝜕4 𝐗(𝑢,𝑣) 𝜕4 𝐗(𝑢,𝑣)
𝑎1 + 𝑎2 = 𝐅(𝑢, 𝑣) (1)
𝜕𝑢4 𝜕𝑣 4

where 𝒂1 and 𝒂2 are called vector-value shape control parameters, which can be used
to change the shape of a PDE surface without changing boundary continuities and each
of which has three components 𝑎𝑖𝑥 , 𝑎𝑖𝑦 , and 𝑎𝑖𝑧 (𝑖 = 1, 2) with 𝒂1 =
[𝑎1𝑥 𝑎1𝑦 𝑎1𝑧 ]𝑇 and 𝒂2 = [𝑎2𝑥 𝑎2𝑦 𝑎2𝑧 ]𝑇 , 𝑢 and 𝑣 are two parametric variables
defined by 0 ≤ 𝑢 ≤ 1 and 0 ≤ 𝑣 ≤ 1, 𝐗(𝑢, 𝑣) is a vector-valued position function
used to define a PDE surface patch, which has three components 𝑥(𝑢, 𝑣), 𝑦(𝑢, 𝑣), and
𝑧(𝑢, 𝑣) with 𝐗(𝑢, 𝑣) = [𝑥(𝑢, 𝑣) 𝑦(𝑢, 𝑣) 𝑧(𝑢, 𝑣)]𝑇 , and 𝐅(𝑢, 𝑣) is a vector-valued
sculpting function, which also has three components 𝑓𝑥 (𝑢, 𝑣) 𝑓𝑦 (𝑢, 𝑣) and 𝑓𝑧 (𝑢, 𝑣)
 5

with 𝐅(𝑢, 𝑣) = [𝑓𝑥 (𝑢, 𝑣) 𝑓𝑦 (𝑢, 𝑣) 𝑓𝑧 (𝑢, 𝑣)]𝑇 . The components of each of the two
vector-valued shape control parameters cab be taken to be the same, i. e, 𝑎𝑖𝑥 = 𝑎𝑖𝑦 = 𝑎𝑖𝑧
(𝑖 = 1, 2) or different, i. e., 𝑎𝑖𝑥 ≠ 𝑎𝑖𝑦 ≠ 𝑎𝑖𝑧 (𝑖 = 1, 2).
In order to simplify mathematical notations, we define the following mathematical
operations in this paper
𝒂1 𝒂2 = [𝑎1𝑥 𝑎2𝑥 𝑎1𝑦 𝑎2𝑦 𝑎1𝑧 𝑎2𝑧 ]𝑇
𝒂1 𝑎1𝑥 𝑎1𝑦 𝑎1𝑧 𝑇
=[ ]
𝒂2 𝑎2𝑥 𝑎2𝑦 𝑎2𝑧
𝑒 𝒂1 = [𝑒 𝑎1𝑥 𝑒 𝑎1𝑦 𝑒 𝑎1𝑧 ]𝑇
𝑇
𝑛𝒂1 𝑛 𝑎1𝑥 𝑛 𝑎1𝑦 𝑛 𝑎1𝑧
√ = [√ √ √ ]
𝒂2 𝑎2𝑥 𝑎2𝑦 𝑎2𝑧
cos𝒂1 = [𝑐𝑜𝑠𝑎1𝑥 𝑐𝑜𝑠𝑎1𝑦 𝑐𝑜𝑠𝑎1𝑧 ]𝑇
sin𝒂1 = [𝑠𝑖𝑛𝑎1𝑥 𝑠𝑖𝑛𝑎1𝑦 𝑠𝑖𝑛𝑎1𝑧 ]𝑇 (2)

In this paper, we look for closed form solutions of the homogeneous form of the
partial differential equation (1) and use one of them for shape reconstruction from point
clouds. In the extended version of this paper, we will investigate the particular solution
of the partial differential equation (1) and use it to develop a more powerful shape re-
construction tool.
We use the method of separation of variables to solve the homogeneous form of the
vector-valued partial differential equation (1) and obtain its four closed form solutions.
The details of solving the homogeneous form of the vector-valued partial differential
equation (1) will be given in the extended version of this paper. In what follows, we
use one closed form solution among the four obtained closed form solutions to demon-
strate shape reconstruction using the closed form solutions of the vector-valued fourth-
order partial differential equation (1) and investigate the errors of shape reconstruction.
The closed form solution to be used can be written in the following form

𝐗(𝑢, 𝑣) = ∑16
𝑗=1 𝒅𝑗 𝒇𝑗 (𝑢, 𝑣) (3)

where
𝒇1 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇2 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇3 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇4 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇5 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 −𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇6 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 −𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇7 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 −𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇8 (𝑢, 𝑣) = 𝑒 𝒒2𝑢 𝑒 −𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇9 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇10 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇11 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇12 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
6

𝒇13 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 −𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣

𝒇14 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 −𝒒4𝑣 𝑐𝑜𝑠𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣
𝒇15 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 −𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑐𝑜𝑠𝒒4 𝑣
𝒇16 (𝑢, 𝑣) = 𝑒 −𝒒2𝑢 𝑒 −𝒒4𝑣 𝑠𝑖𝑛𝒒2 𝑢𝑠𝑖𝑛𝒒4 𝑣 (4)

and

√2 4 𝒄0
𝒒2 = √| |
2 𝒂1
√2 4 𝒄0
𝒒4 = √|𝒂 | (5)
2 2

where 𝒅𝑗 (𝑗 = 1, 2, 3, ⋯ , 16) are the vector-valued unknowns, and 𝒄0 is a vector-val-

ued constant.
𝐗(𝑢, 𝑣) in Eq. (3) defines a PDE surface patch. In the following section, we will
discuss how to use it to achieve shape reconstruction for point clouds.

4 Shape reconstruction and error analysis

Shape reconstruction from point clouds is to find the 16 vector-valued unknowns 𝒅𝑗

(𝑗 = 1, 2, 3, ⋯ , 16) which make the PDE surface patch 𝐗(𝑢, 𝑣) best fit the points in the
region to be reconstructed. For any unorganized point clouds, we can find the points
close to each of a set of planes, which define a curve close to the plane. From these
curves, we can obtain the values of the parametric variable 𝑢 for these curves. From the
points on each of the curves, we can obtain the values of the parametric variable 𝑣 for
all the points on the curve. For the points not on these curves, we can use the geometric
relationships of their positions relative to the points on the curves to obtain the values
of parametric variables 𝑢 and 𝑣. By doing so, the values of the parametric variables 𝑢
and 𝑣 for all the points to be used for shape reconstruction are obtained. That is, for
each point 𝑿𝑛 , we obtain its parametric values 𝑢𝑛 and 𝑣𝑛 .
If 𝑁 points 𝑿𝑛 (𝑛 = 1, 2, 3, ⋯ , 𝑁) are to be used to reconstruct a PDE surface patch
𝐗(𝑢, 𝑣) , we can calculate the squared sum of the errors between the known points 𝑿𝑛
(𝑛 = 1, 2, 3, ⋯ , 𝑁) and the unknown points 𝐗(𝑢𝑛 , 𝑣𝑛 ) with the following equation
𝑁

𝑬 = ∑[𝐗(𝑢𝑛 , 𝑣𝑛 ) − 𝑿𝑛 ]2
𝑛=1
16 2
= ∑𝑁
𝑛=1[∑𝑗=1 𝒅𝑗 𝒇𝑗 (𝑢𝑛 , 𝑣𝑛 ) − 𝑿𝑛 ] (6)

The least squares are used to minimize the error 𝑬 and find the 16 vector-valued
unknowns with the equation below
𝜕𝑬
=0
𝜕𝒅𝑘
(𝑘 = 1, 2, 3, ⋯ , 16) (7)
 7

Substituting Eq. (6) into Eq. (7), we obtain the following equations which can be
used to determine the 16 vector-valued unknowns 𝒅𝑗 (𝑗 = 1, 2, 3, ⋯ , 16)
16 𝑁 𝑁

∑ 𝒅𝑗 ∑ 𝒇𝑗 (𝑢𝑛 , 𝑣𝑛 ) 𝒇𝑘 (𝑢𝑛 , 𝑣𝑛 ) = ∑ 𝑿𝑛 𝒇𝑘 (𝑢𝑛 , 𝑣𝑛 )

𝑗=1 𝑛=1 𝑛=1
(𝑘 = 1, 2, 3, ⋯ , 16) (8)

Equation (8) involves 16 equations which can be used to determine the 16 vector-
valued unknowns 𝒅𝑗 (𝑗 = 1, 2, 3, ⋯ , 16).
𝒇𝑗 (𝑢𝑛 , 𝑣𝑛 ) and 𝒇𝑘 (𝑢𝑛 , 𝑣𝑛 ) in Eq. (8) involve the constants 𝒒2 and 𝒒4 , which are de-
termined by 𝒄0 , 𝒂1 and 𝒂2 , respectively. 𝒂1 and 𝒂2 are vector-valued shape control pa-
rameters. They can be optimized to obtain the optimal PDE surface patch which best
fits to the points 𝑿𝑛 (𝑛 = 1, 2, 3, ⋯ , 𝑁).
However, if we take 𝒒2 and 𝒒4 as design variables, Eq. (8) becomes nonlinear. Solv-
ing Eq. (8) is to solve 16 nonlinear equations which makes the determination of the 16
vector-valued unknowns 𝒅𝑗 (𝑗 = 1, 2, 3, ⋯ , 16) more difficult. Instead of optimizing
𝒒2 and 𝒒4 to find their optimal values, we set 𝒒2 and 𝒒4 to different values, solve the
16 linear algebra equations of Eq. (8), and find that 𝒒2 = 0.1 and 𝒒4 = 0.1 give good
results.
In what follows, we use Eq. (3) to reconstruct a 3D PDE surface from different
points, and compare the surface defined by the original points and the reconstructed
PDE surface. In order to quantify the differences between the two surfaces, we calculate
the maximum error and the average error between the two surfaces with the following
equations
𝐸𝑟𝑟𝑀 = 𝑚𝑎𝑥{|𝑿1 − 𝐗(𝑢1 , 𝑣1 )| |𝑿2 − 𝐗(𝑢2 , 𝑣2 )| ⋯ |𝑿𝑁 − 𝐗(𝑢𝑁 , 𝑣𝑁 )|}
1
𝐸𝑟𝑟𝐴 = ∑𝑁 𝑛=1|𝑿𝑛 − 𝐗(𝑢𝑛 , 𝑣𝑛 )| (9)
𝑁

where 𝐸𝑟𝑟𝑀 indicates the maximum error between the two surfaces, 𝐸𝑟𝑟𝐴 indicates
the average error between the two surfaces, |∙| indicates the distance between the cor-
respondent points of the two surfaces.
Firstly, we consider reconstructing a surface from 16 points give in Table 1. Since a
PDE surface patch (3) involves 16 unknowns, the PDE surface patch should pass
through the 16 points if interpolation operation is used to determine the 16 unknowns.
In this paper, the fitting operation defined by Eq. (8), not interpolation operation, is
used to determine the 16 unknowns. Although the interpolation operation is not used,
it is expected that the fitting operation should give high accuracy if not passing the 16
points.

Table 1. 16 points used to define the surface in Fig. 1.

(𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧)
(0.05, -0.07, -1.12) (0.07, -0.07, -1.11) (0.10, -0.07, -1.11) (0.13, -0.07, -1.12)
(0.04,-0.10, -1.10) (0.07, -0.10, -1.09) (0.10, -0.10, -1.10) (0.13, -0.10, -1.10)
(0.04, -0.13, -1.08) (0.07, -0.13, -1.06) (0.10, -0.13, -1.07) (0.13, -0.13, -1.09)
(0.04, -0.16, -1.05) (0.07, -0.16, -1.04) (0.10, -0.16, -1.04) (0.13, -0.16, -1.06)
8

Setting 𝑁 = 16 in Eq. (8) and solving the 16 linear algebra equations, we obtain the
16 vector-valued unknowns. With the original 16 points, we create the surface depicted
in Fig. 1(a). Substituting the obtained 16 vector-valued unknowns back into Eq. (3), we
use Eq. (3) to create the PDE surface patch shown in Fig. 1(b). The maximum error and
the average error between the two surfaces are given in Table 2.

(a) (b)
Fig. 1. Surface defined by 16 original points and the reconstructed PDE Surface.

Comparing the surface in Fig. 1(a) and the surface in Fig. 1(b), we could not find
any differences, indicating the reconstructed PDE surface is the same as the original
surface defined by the 16 points. This observation is also supported by the maximum
error and average error given in Table 2. The maximum error between the two surfaces
is 5.52 × 10−5 and the average error between the two surfaces is 2.31 × 10−5 . Both
errors are very small, which indicating high accuracy of the fitting operation.

Table 2. Maximum errors and average errors between the two surfaces
𝑁 16 25 36 49 64 81
ErrM 5.52 1.56 2.96 9.68 1.40 2.02
× 10−5 × 10−3 × 10−3 × 10−3 × 10−2 × 10−2
ErrA 2.31 4.47 1.33 3.80 5.76 8.16
× 10−5 × 10−4 × 10−3 × 10−3 × 10−3 × 10−3

Secondly, we consider reconstructing a surface from 25 points consisting of those in

Table 1 and Table 3. Setting 𝑁 = 25 in Eq. (8) and solving the 16 linear algebra equa-
tions, we obtain the 16 vector-valued unknowns. With the original 25 points, we create
the surface depicted in Fig. 2(a). Substituting the obtained 16 vector-valued unknowns
back into Eq. (3), we use Eq. (3) to create the PDE surface shown in Fig. 2(b). The
maximum error and the average error between the two surfaces are given in Table 2.
Comparing the surface in Fig. 2(a) and the surface in Fig. 2(b), we still could not
find any differences, indicating the reconstructed PDE surface is very similar to the
original surface defined by the 25 points. As indicated in Table 2, the maximum error
 9

between the two surfaces is 1.56 × 10−3 and the average error between the two sur-
faces is 4.47 × 10−4 , indicating small errors between the original surface and the re-
constructed PDE surface.

(a) (b)
Fig. 2. Surface defined by 25 original points and the reconstructed PDE surface.

Table 3. Points used with those in Table 1 to define the surface in Fig. 2
(𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧)
(0.02, -0.07, -1.16) (0.02, -0.10, -1.16) (0.07, -0.19, -1.01) (0.10, -0.19, -1.01)
(0.01, -0.19, -1.05) (0.04, -0.19, -1.02) (0.10, -0.10, -1.10) (0.13, -0.10, -1.10)
(0.13, -0.19, -1.03)

Thirdly, we consider reconstructing a surface from 36 points consisting of those in

Table 1, Table 3, and Table 4. Setting 𝑁 = 36 in Eq. (8) and solving the 16 linear al-
gebra equations, we obtain the 16 vector-valued unknowns. With the original 36 points,
we create the surface depicted in Fig. 3(a). Substituting the obtained 16 vector-valued
unknowns back into Eq. (3), we use Eq. (3) to create the PDE surface shown in Fig.
3(b). The maximum error and the average error between the two surfaces are given in
Table 2.

(a) (b)
Fig. 3. Surface defined by 36 original points and the reconstructed PDE surface.
10

Comparing the surface in Fig. 3(a) and the surface in Fig. 3(b), no obvious differ-
ences could be found, indicating the reconstructed PDE surface gives a good approxi-
mation to the original surface defined by the 36 points. As shown in Table 2, the max-
imum error between the two surfaces is 2.96 × 10−3 and the average error between the
two surfaces is 1.33 × 10−3 . Although design variables have been reduced by more
than a half, i. e., from 36 × 3 = 108 to 16 × 3 = 48, the maximum error and average
error are small.

Table 4. Points used with those in Tables 1 and 3 to define the surface in Fig. 3
(𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧) (𝑥, 𝑦, 𝑧)
(0.02, -0.04, -1.20) (0.05, -0.04, -1.15) (0.07, -0.04, -1.14) (0.10, -0.04, -1.14)
(0.13, -0.04, -1.15) (0.17, -0.03, -1.20) (0.17, -0.07, -1.17) (0.16, -0.10, -1.14)
(0.16, -0.13, -1.12) (0.16, -0.16, -1.10) (0.16, -0.19, -1.07)

Fourthly, we consider reconstructing a surface from 49 points. Setting 𝑁 = 49 in

Eq. (8) and solving the 16 linear algebra equations, we obtain the 16 vector-valued
unknowns. With the original 49 points, we create the surface depicted in Fig. 4(a). Sub-
stituting the obtained 16 vector-valued unknowns back into Eq. (3), we use Eq. (3) to
create the PDE surface patch shown in Fig. 4(b). The maximum error and the average
error between the two surfaces are given in Table 2.

(a) (b)
Fig. 4. Surface defined by 49 original points and the reconstructed PDE surface.

The surface in Fig. 4(a) looks like the surface in Fig. 4(b). As indicted in Table 2,
the maximum error between the two surfaces is 9.68 × 10−3 and the average error be-
tween the two surfaces is 3.80 × 10−3 . For this case, the design variables are reduced
by two thirds, i. e., from 49 × 3 = 147 to 16 × 3 = 48, the maximum error and aver-
age error are still small.
Fifthly, we consider reconstructing a surface from 64 points. Setting 𝑁 = 64 in Eq.
(8) and solving the 16 linear algebra equations, we obtain the 16 vector-valued un-
knowns. With the original 64 points, we create the surface depicted in Fig. 5(a). Sub-
stituting the obtained 16 vector-valued unknowns back into Eq. (3), we use Eq. (3) to
 11

create the PDE surface shown in Fig. 5(b). The maximum error and the average error
between the two surfaces are given in Table 2.
The surface in Fig. 5(a) still looks like the surface in Fig. 5(b). The maximum error
between the two surfaces is 1.40 × 10−2 and the average error between the two sur-
faces is 5.76 × 10−3 . For this case, the design variables are reduced by three fourths, i.
e., from 64 × 3 = 192 to 16 × 3 = 48, the maximum error and average error are not
big.

(a) (b)
Fig. 5. Surface defined by 64 original points and the reconstructed PDE surface.

(a) (b)
Fig. 6. Surface defined by 81 original points and the reconstructed PDE surface.

Finally, we consider reconstructing a surface from 81 points. Setting 𝑁 = 81 in Eq.

(8) and solving the 16 linear algebra equations, we obtain the 16 vector-valued un-
knowns. With the original 81 points, we create the surface depicted in Fig. 6(a). Sub-
stituting the obtained 16 vector-valued unknowns back into Eq. (3), we use Eq. (3) to
create the PDE surface shown in Fig. 6(b). The maximum error and the average error
between the two surfaces are given in Table 2.
Since the design variables have been reduced from 81 × 3 = 243 to 16 × 3 = 48,
i. e., reduced by four fifths, some differences can be observed from the two surfaces
depicted in Fig. 6(a) and 6(b). As shown in Table 2, the maximum error between the
12

two surfaces becomes 2.02 × 10−2 and the average error between the two surfaces be-
comes 8.16 × 10−3 , which are not big.

5 Conclusion

In this paper, we have proposed a new method for shape reconstruction from point
clouds. It is based on accurate closed form solutions to a vector-valued fourth-order
partial differential equation. Due to the nature of explicit analytical expressions of the
reconstructed surface, the proposed method is very efficient in reconstructing 3D sur-
faces from point clouds. The error analysis given in this paper demonstrates good ap-
proximation of the reconstructed surface from the proposed PDE-based shape recon-
struction method to point clouds.
There are a number of directions that can be investigated in the following work. First,
the particular solution of the nonhomogeneous equation of Eq. (2) should be developed
to make the proposed PDE-based shape reconstruction more powerful. Second, how to
use the obtained solutions of the proposed vector-valued fourth-order partial differen-
tial equation to reconstruct unorganized point clouds with noise and missing data should
be examined. In this paper, only shape reconstruction by using one single PDE surface
patch was discussed. For complicated 3D models, a single PDE surface patch is unable
to reconstruct complicated 3D models. How to use multiple PDE surface patches to
reconstruct complicated 3D models from point clouds should be considered. The vec-
tor-valued shape control parameters involved in the PDE (2) affects the shape of recon-
structed surfaces. How to identify an efficient and effective optimization method and
combine it with a suitable method of solving the nonlinear equations (8) to find optimal
values of the two parameters 𝒒2 and 𝒒4 , and obtain the 16 vector-valued unknowns to
maximum the potential of the proposed PDE-based shape reconstruction should also be
explored. Another interesting direction is using the obtained closed form solutions to
accurately satisfy all the constraints on four boundaries of a PDE patch. To do this, Eq.
16
(3) can be extended into a series 𝐗(𝑢, 𝑣) = ∑𝑀 𝑚 ∑𝑗=1 𝒅𝑚𝑗 𝒇𝑚𝑗 (𝑢, 𝑣) with 𝒒2 and 𝒒4 be-
ing replaced by 𝒒2𝑚 and 𝒒4𝑚 . We will investigate this in our following work.

Acknowledgements

This research is supported by the PDE-GIR project which has received funding from
the European Union Horizon 2020 research and innovation programme under the Marie
Skodowska-Curie grant agreement No 778035. Zaiping Zhu is also sponsored by China
Scholarship Council.

References
1. Berger, M., Tagliasacchi, A., Seversky, L.M., Alliez, P., Levine, J.A., Sharf, A., Silva, C.T.:
State of the art in surface reconstruction from point clouds. Eurographics 2014 - State of the
Art Reports, pp.161-185, (2014).
 13

2. Berger, M., Tagliasacchi, A., Seversky, L.M., Alliez, P., Guennebaud, G., Levine, J.A.,
Sharf, A., Silva, C.T.: A survey of surface reconstruction from point clouds. Computer
Graphics Forum 36(1), 301-329 (2017).
3. Othman, M.N.M., Yusoff, Y., Haron, H., You, L.H.: An overview of surface reconstruction
using partial differential equation (PDE). IOP Conf. Series: Materials Science and Engineer-
ing 551, 1-5 (2019).
4. Boissonnant, J.D.: Geometric structures for three-dimensional shape reconstruction. ACM
Transactions on Graphics 3(4), 266-289 (1984). Hoppe, H., DeRose, T., Duchamp, T.,
McDonald, J., Stuetzle, W.: Surface reconstruction from unorganized points. In Proceedings
of SIGGRAPH 1992, pp. 71-78 (1992).
5. Hoppe, H., DeRose, T., Duchamp, T., McDonald, J., Stuetzle, W.: Surface reconstruction
from unorganized points. In Proceedings of SIGGRAPH 1992, pp. 71-78 (1992).
6. Oblonšek, Č., Guid, N.: A fast surface-based procedure for object reconstruction from 3D
scattered points. Computer Vision and Image Understanding 69(2), 185-195 (1998).
7. Bernardini, F., Mittelman, J., Rushmeier, H., Silva, C., Taubin, G.: The ball-pivoting algo-
rithm for surface reconstruction. IEEE Transactions on Visualization and Computer
Graphics 5(4), 349-359 (1999).
8. Gopi, M., Krisnan, S., Silva, C.: Surface reconstruction based on lower dimensional local-
ized Delaunay triangulation. Computer Graphics Forum 19(3), 467-478 (2000).
9. Lee, A., Moreton, H., Hoppe, H.: 2000: Displaced subdivision surfaces. Proceedings of
SIGGRAPH 2000, pp. 85-94 (2000).
10. Jeong, W.K., Kim, C.H.: Direct reconstruction of displaced subdivision surface from unor-
ganized points. Graphical Models 64(2), 78-93 (2002).
11. Nan, L., Wonka, P.: PolyFit: Polygonal surface reconstruction from point clouds. 2017 IEEE
International Conference on Computer Vision (ICCV), Venice, pp. 2372-2380 (2017).
12. Duan, Y., Yang, L., Qin, H., Samaras, D.: Shape reconstruction from 3D and 2D data using
PDE-based deformable surfaces. Lecture Notes in Computer Science 3023, 238-251 (2004).
13. Linz, C., Goldlücke, B., Magnor, M.: A point-based approach to PDE-based surface recon-
struction. Lecture Notes in Computer Science 4174, 729-738 (2006).
14. Franchini, E., Morigi, S., Sgallari, F.: Implicit shape reconstruction of unorganized points
using PDE-based deformable 3D manifolds. Numerical Mathematics: Theory, Methods and
Applications 3(4), 405-430 (2010).
15. Pana, R., Skala, V.: Continuous global optimization in surface reconstruction from an ori-
ented point cloud. Computer-Aided Design 43, 896-901 (2011).
16. Liu, W., Cheung, Y., Sabouri, P., Arai, T.J., Sawant, A., Ruan, D.: A continuous surface
reconstruction method on point cloud captured from a 3D surface photogrammetry system.
Medical Physics 42(11), 6564-6571 (2015).
17. He, Y., Huska, M., Kang, S.H., Liu, H.: Fast algorithms for surface reconstruction from
point cloud, https://arxiv.org/abs/1907.01142, pp. 1-16 (2019).
18. Li, J., Hero, A.O.: A fast spectral method for active 3D shape reconstruction. Journal of
Mathematical Imaging and Vision 20, 73-87 (2004).
19. Ugail, H., Kirmani, S.: Method of surface reconstruction using partial differential equations.
Proceedings of the 10th WSEAS International Conference on COMPUTERS, Vouliagmeni,
Athens, Greece, July 13-15, pp. pp51-56 (2006).
20. Elyan, E., Ugail, H.: Reconstruction of 3D human facial images using partial differential
equations. Journal of Computers 2 (8): 1-8 (2007).
21. Rodrigues, M., Osman, A., Robinson, A.: Partial differential equations for 3D data compres-
sion and reconstruction. Advances in Dynamical Systems and Applications 8(2): 303-315
(2013).
14

22. Sheng, B., Zhao, F., Yin, X., Zhang, C., Wang, H., Huang, P.: A lightweight surface recon-
struction method for online 3D scanning point cloud data oriented toward 3D printing. Math-
ematical Problems in Engineering 2018, Article ID 4673849: 1-16 (2018).
23. Iglesias, A., Echevarrí a, G., Gálvez, A.: Functional networks for B-spline surface recon-
struction. Future Generation Computer Systems 20(8), 1337-1353 (2004).
24. Gálvez, A., Iglesias, A.: Iterative two-step genetic-algorithm-based method for efficient pol-
ynomial B-spline surface reconstruction. Information Sciences 182(1), 56-76 (2012).
25. Gálvez, A., Iglesias, A.: Particle swarm optimization for non-uniform rational B-spline sur-
face reconstruction from clouds of 3D data points. Information Sciences 192(1), 174-192
(2012).