Point Cloud Neighborhood Estimation Method Using Deep
Neuro-Evolution
Ahmed Abouelazm, Igor Vozniak, Nils Lipp, Pavel Astreika and Christian Mueller
Deutsches Forschungszentrum f
¨
ur K
¨
unstliche Intelligenz (DFKI), Saarbruecken, Germany
Keywords:
Point Clouds, 3D Deep Learning, Neighborhood Estimation, Deep Neuro-Evolution.
Abstract:
Due to the recent advancements in computing hardware, deep learning, and 3D sensors, point clouds have
become an essential 3D data structure, and their processing and analysis have received considerable attention.
Given the unstructured and irregular nature of point clouds, encoding local geometries is a significant barrier in
point cloud analysis. The aforementioned challenge is known as neighborhood estimation, and it is commonly
addressed by fitting a plane to points within a local neighborhood defined by estimated parameters. The
estimated neighborhood parameters for each point should adapt to the point cloud’s irregularities and different
local geometries’ sizes and shapes. Different objective functions have been derived in the literature for optimal
parameters selection with no efficient approach for these objective functions’ optimization as of now. In this
work, we propose a novel neighborhood estimation pipeline for such optimization which is objective function
and neighborhood type invariant, utilizing a modified version of deep Neuro-Evolution algorithm and Farthest
Point Sampling as an intelligent sampling approach. Results demonstrate the ability of the proposed pipeline
for state-of-the-art objective functions optimization and enhancement of neighborhood properties estimation
such as the normal vector.
1 INTRODUCTION
RGB images have long been the focus of image pro-
cessing and computer vision research. Furthermore,
the fast development of deep learning has aided the
rise of 2D images as a dominant data structure (Chai
et al., 2021). However, we live in a three-dimensional
world, and the capacity of two-dimensional images to
fully depict three-dimensional objects in the world is
limited. This is due to the fact that RGB images do not
capture depth information and instead project data on
a single depth 2D plane. Furthermore, RGB images
are sensitive to illumination and lack contrast. Due
to the importance of depth information in understand-
ing the geometry of a scene, algorithmic techniques
for depth estimation from single and stereo images
have been investigated in the literature (Ming et al.,
2021; Lazaros et al., 2008). Despite the success of
these techniques, 3D sensors that can directly capture
depth without estimation are required. Thus, several
3D data formats have been proposed in recent years,
such as point clouds and RGB-D images. Due to the
rapid development of 3D acquisition devices such as
LiDAR and RGB-D sensors, as well as the decrease
in their cost, 3D data structures have been expanded
to a variety of applications nowadays such as robotics
and autonomous driving (Royo and Ballesta-Garcia,
2019). One of the most prevalent 3D data structures
Figure 1: Challenge of neighborhood estimation on an ir-
regular point cloud (Bello et al., 2020), as the point cloud
has sparse and dense regions, which requires adapting
neighborhood parameters.
is Point Cloud Data (PCD), which is a representation
generated by LiDAR sensors. PCD is a 3D format
that depicts a scanned object as a set of discrete points
scattered in a Euclidean space. Each point representa-
tion can be augmented by color, normal vector at the
point of interest, and other feature descriptors. De-
spite becoming increasingly popular in the research
community, PCDs still face the following main dis-
advantages which must be considered during analysis
and processing. PCD is an unordered set of points that
lacks any underlying structure or regular grid. Be-
sides, it is an irregular data type with some regions
having high density, large number of points per unit
582
Abouelazm, A., Vozniak, I., Lipp, N., Astreika, P. and Mueller, C.
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution.
DOI: 10.5220/0011627300003417
In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages
582-593
ISBN: 978-989-758-634-7; ISSN: 2184-4321
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
volume, whereas others are sparse, with just a few
points as depicted on figure 1.
The challenges as mentioned earlier require per-
mutation invariant density aware methods for encod-
ing local geometries and local correlations between
points in a PCD. Local geometries are commonly en-
coded by extracting features from a plane fitted to a
neighborhood, where the points included in the neigh-
borhood are determined by a set of parameters, for
each point in the PCD. The estimation of these pa-
rameters is known as neighborhood estimation (Wein-
mann et al., 2015; Demantk
´
e et al., 2011), and it
is a critical step for the majority of machine learn-
ing techniques used in PCD processing. Since the
attributes initially retrieved using neighborhood es-
timation can be used for more complex tasks such
as keypoints detection and feature descriptors (Ciru-
jeda et al., 2014; Huang and You, 2012; Siritanawan
et al., 2017; Zhong, 2009). Furthermore, neighbor-
hood estimation is not only essential for machine
learning techniques but has also been integrated into
deep learning modules (Liu et al., 2019; Suo et al.,
2020; Zhao et al., 2022; Yang et al., 2018).
Due to the obvious irregular nature of PCDs, as
illustrated on figure 1, fitting a neighborhood with the
same parameters throughout the PCD is inappropriate
as these parameters should vary between dense and
sparse regions and take into account the dimensional-
ity and size of local geometries present in the PCD.
In the literature, several neighborhood types and two
objective functions have been proposed for optimal
neighborhood estimation (Weinmann et al., 2015; De-
mantk
´
e et al., 2011). However, to the best of our
knowledge, no effective pipeline exists to optimize
these suggested objective functions, except for a grid-
based search with fixed resolution presented in (De-
mantk
´
e et al., 2011; Weinmann et al., 2015). The pro-
posed grid search requires exhaustive iteration over a
discrete set of neighborhood parameters fixed for all
points, and doesn’t have the ability to generalize to
similar points.
The contribution of this work is the design
and validation of a novel and computationally effi-
cient (requires fewer training parameters; no back-
propagation step is involved, thus, less used mem-
ory) pipeline for explicit and optimal neighborhood
estimation using a learning module in an unsuper-
vised manner. The learning module learns optimal
parameter selection from a small subset of points and
generalizes the learned selection across the remain-
ing points in the PCD, thus enhancing computation
and memory efficiency while finding an accurate solu-
tion. An effective subset of points for learning is sam-
pled using Farthest Point Sampling (FPS) (Eldar et al.,
1994) as it has the ability to sample points that provide
sufficient information about the point cloud structure.
The novelty of this work is twofold: a) we propose
to use a modified version of Deep Neuro-Evolution
(Salimans et al., 2017) as a learning module such that
parameter selection can be learned over the sampled
points, and the learned selection can generalize for
unseen points that have similar points distribution in
their vicinity. We employ Deep Neuro-Evolution as
it is able to handle non-differentiable objectives such
neighborhood estimation with minimal memory and
computation loads. b) Additionally, we propose, a
novel heuristic function to guide the learning module
search, rejecting trivial solutions of empty neighbor-
hoods.
2 RELATED WORK
The standard approach for encoding local geometries
in PCDs is fitting a plane to a neighborhood cen-
tered at the point of interest. The approximate plane
eigenvalues and eigenvectors are used to estimate key
properties of the interest point such as normal vector,
principle curvature and surface variation (Weinmann
et al., 2014). As a result, the neighborhood estimation
procedure is a critical component of a number of PCD
applications such as keypoints detection and features
description and have been integrated into deep learn-
ing modules. Figure (2) illustrates neighborhood es-
timation and its various applications. In this section,
we provide an overview of related work, relevant for
the objectives of this work.
2.1 Neighborhood Types
There are two types of neighborhoods commonly
used in PCDs analysis: spatial-based and similarity-
based neighborhoods, as categorized in (Weinmann
et al., 2014). Spatial neighborhoods assign points to a
neighborhood v
r
p
centered around the center point p,
if the points are located in the interior surface of a pa-
rameterized geometry representing the neighborhood.
The spherical neighborhood is the most commonly
used 3D spatial neighborhood. Spherical neighbor-
hood fits a sphere parameterized by radius r to a cen-
ter point p, and a point q belongs to the neighbor-
hood of p when the Euclidean distance between them
is smaller than the radius, as shown in equation (1).
Cylindrical neighborhood is another variant of spatial
based neighborhoods, which centers an infinitely long
vertically oriented cylinder with a radius r at a center
point p, as discussed in equation (2) with the height as
the z direction. This neighborhood can be conceived
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
583
Figure 2: Point cloud neighborhood estimation and its vari-
ous applicability cases.
as the projection of points onto a plane with a constant
height and the assignment of points to the neighbor-
hood that are located within a circle of radius r.
r R, v
r
p
=
{
q | d(q, p) r
}
(1)
r R, v
r
p
=
q |
(p
x
q
x
)
2
+ (p
y
q
y
)
2
1
2
r
(2)
K-nearest neighborhoods is the most commonly uti-
lized similarity neighborhood. The k-nearest neigh-
borhood computes a similarity metric over all points
in the PCD. The top-k similar points are included in
the final neighborhood. The Radial Basis Function
(RBF) kernel with a unit variance of the Euclidean
distance between the center point p and the surround-
ing points q is a regularly used similarity metric.
2.2 Plane Fitting
Following the selection of a neighborhood type and
parameters, relevant information is extracted from the
covariance matrix computed over the neighborhood’s
points. As described in equation (3), the covariance
matrix is a symmetric positive semi-definite matrix
computed using statistics over points within a neigh-
borhood v
r
p
centered at point p and parameterized by
r. The eigenvalues and eigenvectors of the covari-
ance matrix, recovered via eigenvalue decomposition
as seen in equation (4), are used to approximate the
fitted plane.
Cov(v
r
p
) =
1
|v
r
p
| 1
qv
r
p
(q µ)
T
× (q µ) (3)
λ ×Cov(v
r
p
) = λ × v (4)
2.3 Keypoints Detection
The majority of keypoints detection algorithms are
dependent on neighborhood estimation as an initial
processing step. Keypoints detection extracts points
with unique features within a local vicinity in a PCD.
ISS (Zhong, 2009) defines keypoints as points with
high 3D variation (smallest eigenvalue of the approx-
imate plane) within a local vicinity. To eliminate
any ambiguity in keypoints, ISS employs two pruning
techniques: points with equal eigenvalues omission
and resolution control. SURE (Fiolka et al., 2012) in-
vestigates the ratio between the approximated plane’s
maximum and minimum eigenvalues. Keypoints are
detected by applying a threshold operator on the in-
vestigated ratio.
MSV (Siritanawan et al., 2017) and MoPC
(Huang and You, 2012) apply multi-scale analysis by
fitting neighborhoods of different scales to a point
of interest. The scales range between minimum and
maximum scales [s
min
, s
max
] and are increased with a
discrete step size. MSV keypoints are points whose
surface variation (smallest normalized eigenvalue of
the approximate plane) averaged over all scales is
higher than a certain threshold. MoPC, on the other
hand, selects keypoints based on principle curvature
(largest eigenvalue of the approximate plane). A point
at a given scale s
n
is considered a keypoint if it has the
largest principle curvature, when compared to both,
the same point at the previous scale s
n1
and the sub-
sequent scale s
n+1
, and all its neighbors that lie within
a third of the scale
1
3
s
n
of the point.
2.4 Features Description
The detection of keypoints is followed by a descrip-
tion of their features in the encoding pipeline. The
aim of features description is to provide an extended
features vector for a more meaningful embedding of
PCD. (Vandapel et al., 2004) proposed a linear com-
bination of the eigenvalues of the estimated plane of
the keypoint, yielding an eigenvalue-based descriptor
which is resilient against rigid transformations. RSD
(Marton et al., 2010) models a radial relationship be-
tween a keypoint and its neighbors by the distance be-
tween them and the angle between their normals. The
final descriptor is minimum and maximum radius at-
tained within a neighborhood. (Triebel et al., 2006)
utilized a 1D histogram of cosine angles between a
keypoint and its neighborhoods normal vectors as a
descriptor. Thrift (Flint et al., 2007) is another ap-
proach that uses 1D histograms of cosine angles, with
the angle defined between normal vectors computed
on two different scales for the same keypoint neigh-
bor.
Spin Image (Johnson and Hebert, 1998) calculates
two geometric features (α, β) between a keypoint p
and its neighbor q by aid of the neighbor’s normal
vector n
q
. A 2D histogram over α, β of the keypoint
neighborhood is the proposed descriptor. The ap-
proach of PFH (Rusu et al., 2008b; Rusu et al., 2008a)
is similar to that of Spin Image. Given a keypoint p
and a pair of its neighbors, q
s
and q
t
, a darboux frame
is constructed based on the distance between the pair
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
584
and the normal vector of q
s
. Four geometric proper-
ties are derived using the darboux frame, and the final
PFH descriptor is a 4D histogram that incorporates
these properties.
2.5 Deep Learning
The neighborhood estimation task has been used in a
number of deep learning modules. PointNet (Qi et al.,
2017a) is a pioneer work in deep learning on the raw
data of PCDs. Even though this network provided ad-
vancement in the field and sparked huge interest in
learning from raw PCD, it lacked the ability to cap-
ture the correlation between neighboring points and
complex local geometries. Thus, (Qi et al., 2017b)
proposed PointNet++ to overcome this drawback by
hierarchical learning contextual information over a
PCD. PointNet++ samples a subset of points from a
PCD and groups points in a fixed radius, fixed neigh-
borhood parameter, around the sampled points using
a sampling and grouping layer. The grouped points
in the local neighborhood are processed using Point-
Net to extract useful information. DGCNN (Wang
et al., 2019) constructs a k-nearest neighbors graph
with fixed K and uses an edge convolution layer to
learn from its constructed input. FoldingNet (Yang
et al., 2018) is a PCD auto-encoder architecture used
for reconstruction task. This architecture appends the
local covariance matrix of each point, calculated us-
ing a knn-neighborhood with K=16 as described in
equation (3), to the input feature vector. Both LPD-
Net (Liu et al., 2019) and its successor LPD-AE (Suo
et al., 2020) merge ten features that were retrieved
from the fitted neighborhood with the input points
coordinates. The empirical results in the aforemen-
tioned papers and their analysis showed that select-
ing the optimal neighborhood size rather than a fixed
one improved network performance and recall abil-
ity. However, a grid search with fixed resolution
over neighborhood parameters was used, which has
the drawback of requiring a large amount of memory
and computing load. (Zhao et al., 2022) dedicated a
separate branch to extracting information from fea-
tures computed over a knn-neighborhoods with con-
stant size for all points. The retrieved features are cal-
culated using local frames constructed between a cen-
ter point p and its neighbor q, and are then fused with
extracted features from spatial information to form a
single descriptor. The addition of such features in-
creased the network’s robustness against rigid input
transformations.
2.6 Neighborhood Estimation Objective
Functions
(Demantk
´
e et al., 2011) identifies two major issues
with selecting neighborhood parameters: it cannot be
chosen heuristically, and it cannot be constant over
the whole PCD. Instead, it should be directed by data
and tailored to varying densities and sizes of PCD lo-
cal geometries.
(Demantk
´
e et al., 2011) introduces two objective
functions for selecting neighborhood parameters. The
goal of such functions is to automate the selection of
the optimal neighborhood parameters for each point
in a PCD. Both objective functions are based on a
point’s dimensionality features, as described in (To-
shev et al., 2010). These features are calculated based
on eigenvalues ordered in descending order as de-
picted in equation (5), where a
1D
, a
2D
, and a
3D
rep-
resent the degree of linearity, planarity, and scattering
exhibited by a point p when a neighborhood v param-
eterized by r is centered around p. The dimension-
ality label of a point p is the degree with the highest
value as provided in equation (6), where λ
1
λ
2
λ
3
represent the eigenvalues of the approximate plane in
descending order.
a
1D
=
λ
1
λ
2
λ
1
, a
2D
=
λ
2
λ
3
λ
1
, a
3D
=
λ
3
λ
1
(5)
d
(v
r
p
) = argmax
d[1,3]
a
dD
(6)
The first objective function defines the optimal pa-
rameters as the parameters that minimize the Shannon
entropy of the dimensionality features within a neigh-
borhood, as described in equation (7, 8). The intuition
behind such an objective function is finding parame-
ters which favors a single dimensionality within the
neighborhood.
E(v
r
p
) = a
1D
ln(a
1D
) a
2D
ln(a
2D
) a
3D
ln(a
3D
)
(7)
r
optimal
= arg min
r
E(v
r
p
) (8)
The other objective function supports picking a
neighborhood with the majority of points labeled with
the same dimensionality (homogeneous), as shown in
equations (9, 10).
S(v
r
p
) =
1
n
qv
r
p
1
{
d
(v
r
p
)=d
(v
r
q
)
}
(9)
r
optimal
= arg max
r
S(v
r
p
) (10)
In the aforementioned work, both objective func-
tions were optimized using grid search with a prede-
fined resolution bounded between (r
min
, r
max
). When
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
585
compared to homogeneity-based criteria, the reported
results show that dimensionality-based entropy crite-
ria are more desirable since they can have a unique
solution per point and do not saturate.
An alternative entropy based objective function
is introduced in (Weinmann et al., 2015). Unlike
dimensionality-based functions, this function seeks to
reduce the Shannon entropy of the normalized eigen-
values, as proposed in equations (11, 12). The idea
behind such a function is to decrease noise and uncer-
tainty in the selected neighborhood eigenvalues.
ε
i
=
λ
i
λ
1
+ λ
2
+ λ
3
(11)
E(v
r
p
) = ε
1
ln(ε
1
) ε
2
ln(ε
2
) ε
3
ln(ε
3
) (12)
3 METHODOLOGY
As discussed in Section 2 and illustrated in figure (2),
explicit neighborhood estimation is an essential pre-
processing step for efficient PCD analysis. In this
work, we propose a pipeline for explicit and opti-
mal neighborhood estimation. The pipeline doesn’t
require supervision or access to ground truth infor-
mation such as normal vectors or curvatures. It is
worth mentioning that the estimated neighborhoods
are task-agnostic, with desirable properties due to the
choice of the objective function. Additionally, the
neighborhood is designed with low memory and com-
putation footprint to accelerate the task significantly.
A pipeline for neighborhood estimation that is
both memory and computationally efficient is lacking
in the literature, since only grid search with a fixed
resolution was used to minimize objective functions
addressed in Section 2.6. The two objective functions
depicted in equations (7, 12) are used for automated
neighborhood estimation in this work. The third ob-
jective function presented in equation (9) is disre-
garded as it has a combinatorial nature, and empirical
results reported in (Demantk
´
e et al., 2011) demon-
strate that it performs poorly when compared to the
other two. In this work, we focus on increasing the
task efficiency by learning the optimal neighborhood
parameter choice over a small subset of points from
a PCD driven by the aforementioned objective func-
tions and then generalizing this learning across the re-
maining points.
This work adopts and modifies Deep Neuro-
Evolution (Salimans et al., 2017) as the building block
of the proposed pipeline. To the best of our knowl-
edge, this work is the first in modifying Deep Neuro-
Evolution as a parameter estimation module. Deep
Neuro-Evolution (Such et al., 2017) is a non-gradient-
based alternative to deep reinforcement learning and
is described as the successor to evolution strategies in-
troduced in (Salimans et al., 2017). Evolution Strate-
gies (ES) is a gradient-based deep evolution technique
that produces results comparable to Q-learning and
policy gradient approaches (Salimans et al., 2017).
ES evaluates a network at a number of mutations gen-
erated by random noise. A gradient approximated
by a modified finite-difference operation is used to
update the network parameters. The success of ES
sparked interest in developing an entirely gradient-
free technique based on evolutionary algorithms that
can scale effectively as a deep reinforcement learn-
ing alternative, leading to the formulation of Deep
Neuro-Evolution. In comparison to deep reinforce-
ment learning and ES, Deep Neuro-Evolution has
a sizable advantage in the neighborhood estimation
problem and requires far less memory compared to
reinforcement learning because it does not require
state representation or buffer memory. Additionally,
since Deep Neuro-Evolution does not require back-
propagation, it requires less computational effort than
approaches as in (Such et al., 2017). Furthermore, it
can also be used in a distributed manner over multiple
threads.
Preliminaries. Deep Neuro-Evolution is a variant
of the standard genetic algorithm, in which each in-
dividual (genotype) is a neural network. A mutation
function ψ, network initialization method φ, and fit-
ness function F are all required by the algorithm in
addition to the number of individuals in each gener-
ation, the number of generations, and the number of
top T performing individuals to keep for next gen-
erations. The mutation function ψ modifies the net-
work’s weights to generate a new individual, while
the network initialization φ provides a deterministic
method for creating the initial network’s weights. Fi-
nally, the fitness function is utilized to assess each in-
dividual’s performance.
Our proposed Deep Neuro-Evolution uses a shal-
low version of PointNet (Qi et al., 2017a) as an indi-
vidual, which uses blocks of Multi-Layer Perceptron
(MLP) and Relu Nonlinearity to process each point
in the input individually. In this work, the shallow
network has two PointNet blocks, thus having very
few learnable parameters, decreasing the computation
load. The second block’s output is collapsed by max
pooling, as it is a permutation invariant function due
to its symmetric nature, into a global description. A
linear layer is added after the global description vec-
tor to obtain the network’s final output. The network
input is the PCD centered around the point of interest
p for which the neighborhood is estimated. More-
over, the network computation can be enhanced by
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
586
Figure 3: Pipeline overview. A subset of points is sampled using FPS, which are used to train the proposed pipeline. The
proposed pipeline searches for the Elite Networks for the neighborhood estimation task. The Elite Networks are subsequently
used to estimate parameters of the remaining points.
only taking a sufficient subset of the PCD around the
point of interest. The network output is the neigh-
borhood parameterization and is subjected to a tanh
non-linearity to constrain it to the range [-1, 1], after
which it is remapped to fulfil the chosen parameter
limits.
Deep Neuro-Evolution searches for the fittest in-
dividual (Elite) among those created over G genera-
tions. The fitness of each individual in estimating a
neighborhood for a center point p is evaluated using
our novel fitness function proposed in equation (13).
The fitness function’s first term is in charge of pro-
moting neighborhood parameter selection that mini-
mizes neighborhood entropy and represents the dif-
ference between the maximum Shannon entropy of
a neighborhood defined over three random variables
(x, y, z), which is equal to ln (3), and the neighborhood
entropy obtained with network estimated parameters
r
i
. As a consequence, the fittest parameter choice is
the one with the lowest entropy, as required by equa-
tion (8). Furthermore, this term is independent of the
entropy function employed, and it can be calculated
using either the eigen entropy or the dimensionality
entropy stated in equations (7, 12). The second term
of equation (13) is responsible for penalizing neigh-
borhoods with less than two points in the neighbor-
hood, since this is a trivial solution and doesn’t pro-
vide any tractable information about the center point
p.
F(θ
i
) =
ln(3) E(v
r
i
p
)
+ ln(3) × 1
|
v
r
i
p
|
>2
(13)
In the initial generation, N 1 individuals are
created with randomly initialized weights using the
Xavier uniform initialization (Glorot and Bengio,
2010). The benefit of this initialization is that the net-
work is not biased towards any single region in the
search space, but instead has a uniform distribution of
output over the whole search space.
For subsequent generations, one of the top T -
performing parents is uniformly sampled at random.
The sampled parent weights are mutated by applying
an additive Gaussian noise regulated by a hyperpa-
rameter σ as seen in equation (14), to generate a new
individual which explores a slightly different region
in the search space. The hyperparameter σ controls
how dispersed the new individual is from its parent
in the search space. A high sigma value causes sig-
nificant jumps in the search process, whereas a low
sigma causes the search to be localized and limited to
a certain region of the search space. As a result, the
initial value of σ is obtained empirically during exper-
iments, and σ decays at a linear rate every generation,
allowing the search to be more localized in the fittest
areas of the search space as the number of generations
increases. This concept is parallel to the exploration-
exploitation trade-off in reinforcement learning (Sali-
mans et al., 2017).
ε N (0,I)
θ
n
= ψ (θ
n1
, ε) = θ
n1
+ σ × ε
(14)
Each mutated individual performance is evaluated
using fitness function F. Consequently, the generated
individuals are ordered in a descending manner based
on their fitness score. The set of elite candidates is es-
tablished through truncation selection, which involves
picking the top T -performing parents. The elite indi-
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
587
vidual, the best performing one globally during the
search process thus far, is updated by comparison to
the maximum argument over the candidate set and ap-
pended to the candidate set if not already a member.
The elite individual (network) with the fittest neigh-
borhood parameters is returned as the algorithm out-
put once all generations have been generated. This
elite individual can be utilized in estimating neighbor-
hood parameters for other points that exhibit similar
geometric and spatial properties.
Even though the adapted Deep Neuro-evolution
approach has low memory and processing, it is in-
efficient for estimating neighborhood parameters per
point, especially on large scale PCDs with millions
of points. It is more efficient to learn the parame-
ter search using Deep Neuro-evolution over a small
sample of points, 5% of the points in the PCD, and
then applying the learned search over the remaining
95% unseen points in the PCD. The overall pipeline
pseudocode is illustrated in algorithm 1. It should be
highlighted that pipeline efficiency in both the Deep
Neuro-evolution (lines 2–5) and the elite networks’
evaluation (lines 6–14) stages are further improved by
utilizing a relevant subset of points as the network in-
put instead of the whole PCD e.g. the points included
within the neighborhood defined by the maximum pa-
rameter. Additionally, a visual clarification of the pro-
posed pipeline can be seen in figure (3).
Different sampling techniques have been pro-
posed in the literature to extract keypoints from a PCD
as discussed in Section 2.3. However, since most
of these approaches rely on neighborhood estimation,
using them in the proposed pipeline is contradictory.
Whereas, random sampling and FPS were described
as effective keypoints sampling alternatives for PCD
in (Varga et al., 2021; Qi et al., 2017b) respectively.
Random sampling draws a subset of points from a
PCD uniformly at random. FPS, on the other hand,
is an iterative sampling technique that starts with a
random point as a seed and samples the PCD’s far-
thest point from the previously sampled seed as the
new seed in an iterative manner. In this work, FPS
is preferable since it captures the PCD structure and
summarizes its details in a few points, whereas ran-
dom sampling loses shape details and rough struc-
ture, and can easily hinder the pipeline performance,
as shown by the example on figure (4c).
Classical multi-scale shape analysis (Demantk
´
e
et al., 2011) proposes multiple analysis scales fixed
for all points, which are not tailored by the data. Our
pipeline is more effective than multi-scale analysis
since it requires less computational power and di-
rectly extracts a single optimal value for each param-
eter, which can then be used for the precise estima-
Algorithm 1: The proposed pipeline Pseudocode.
Require: mutation function ψ, population size N,
number of selected individuals T , number of gen-
erations G, network initialization routine φ, fit-
ness function F, point cloud P
1: sample 5% of the total points in P as keypoints Q
using FPS as in line with (Qi et al., 2017b)
2: for each keypoint q in Q do
3: elite Deep Neuro-Evolution(ψ, N, T , G, φ,
F, P centered around q) as in line with (Such
et al., 2017)
4: Elite networks Elite networks elite
5: end for
6: for each remaining point p in P \ Q do
7: for each elite in Elite networks do
8: neighborhood parameters elite (P cen-
tered around p)
9: neighborhood fitness F (neighborhood
parameters)
10: if best fitness < neighborhood fitness then
11: best fitness neighborhood fitness
12: optimal parameters neighborhood pa-
rameters
13: end if
14: end for
15: Per-point parameters Per-point parameters
optimal parameters
16: end for
17: return Per-point parameters
tion of a center point characteristics like normal vec-
tors and curvatures and employed in clustering frame-
works like the region growing algorithm (Rabbani
et al., 2006). Furthermore, our pipeline is easily ex-
tensible as an enhanced multi-scale analysis approach
by altering the pipeline’s output to be the top k
elite individuals for each point in the keypoints, hence
strengthening the multi-scale analysis by selecting in-
formative scales based on geometric and spatial prop-
erties of points rather than a pre-defined set identical
for all points.
It is essential to emphasize, that the proposed
pipeline is an unsupervised approach guided by a
novel objective function for hard assignment of lo-
cal neighborhoods. There are other approaches that
focus on soft assignment of neighborhoods, such as
(Ben-Shabat et al., 2019; Zhou et al., 2021a; Ben-
Shabat and Gould, 2020; Zhu et al., 2021; Zhou et al.,
2021b). The soft assignment is established for a cen-
ter point by learning weights over neighboring points
on a rather big neighborhood size fixed for all cen-
ter points. However, the aforementioned approaches
are trained in a supervised manner to minimize nor-
mal estimation loss given its ground truth and uti-
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
588
(a) Input point cloud.
.
(b) Output point cloud using FPS.
.
(c) Output point cloud using random
sampling.
Figure 4: Output point clouds of a chair class from different sampling techniques. The output illustrates that FPS is favorable
than random sampling as it has the ability to summaries a point cloud in few points while maintaining its structure.
lize a large patch of points around a center point
without explicit extraction of neighborhood parame-
ters. Whereas, our method explicitly extracts neigh-
borhood parameters and is task-agonistic such that
the neighborhood has desired properties that can be
later fine-tuned for specific tasks such as the ones dis-
cussed in Section 2, which are out of the scope of this
research. Moreover, in (Zhou et al., 2021b), one of
the best-performing networks on supervised normal
estimation, authors reported that utilization of a first-
order jet fitting, equivalent to the plane fitting used
in our work, and top k selection strategy enhances
normal estimation results. The top k selection only
uses information from the top k critical points for
fitting, however, it is manually set and fixed for all
points. In our paper, we provided a method for the
dynamic setting of the k parameter of an unweighted
fitting approach. Furthermore, our work can directly
benefit pre-processing steps of other approaches such
as (Yang et al., 2018; Suo et al., 2020; Liu et al.,
2019), which the aforementioned approaches cannot
contribute to. Finally, the proposed approach has con-
siderably fewer learnable parameters than all other
networks and can be used to estimate parameters for a
single PCD without any pre-training. For future work,
it is planned to integrate both jet fitting and weighted
neighborhoods in our pipeline while keeping its unsu-
pervised setting.
4 EVALUATION
4.1 Dataset
In this work, ModelNet-10 dataset is used for the eval-
uation purposes of the proposed neighborhood esti-
mation pipeline. ModelNet-10 is a benchmark for
3D object classification and retrieval published by
Princeton University in (Wu et al., 2015). The dataset
contains, 4899 CAD-generated meshes stored in Ob-
ject File Format (OFF), which are divided into 3991
meshes for training and 908 meshes for testing pur-
poses. Moreover, it is well known in the research
community since it is a well-structured dataset with
Figure 5: Different examples of points irregularity and com-
plex features that require the adaptation of the neighbor-
hood parameters.
pre-aligned clean shapes sampled from several cate-
gories. Currently, the dataset has ten classes: bathtub,
bed, chair, desk, dresser, monitor, nightstand, sofa,
table, and toilet. An output PCD is generated by
randomly sampling points from the CAD-generated
mesh’s triangular faces, which enforces the existence
of irregularities in the PCD. The sampled points are
normalized and fitted into a bounding box between
[1, 1]. Even though ModelNet 10 dataset is a clean
and aligned dataset, it contains a variety of objects
with complex geometries that necessitate the adapta-
tion of the neighborhood parameters. Figure (5) illus-
trates an example of a PCD from class chair. Different
regions with less dense points distribution and irreg-
ular nature are highlighted in figure (5: sub-figures
2 and 5). Additionally, these examples clarify that
edges, corners and other fine details require adapta-
tion of the neighborhood parameters to be captured
correctly compared to points belonging to coarser fea-
tures such as the seat or the back of the chair as de-
picted in figure (5: sub-figures 3, 1, 4 respectively).
4.2 Evaluation Metrics
A k-nearest neighborhood using RBF kernel of Eu-
clidean distance as a similarity metric and constant
size K = 30 is utilized as a baseline against which
this work was evaluated, which is the standard ap-
proach for fitting a neighborhood to the ModelNet-
10 dataset (Qi et al., 2017b; Wang et al., 2019). The
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
589
primary evaluation metrics in this work are: a) the
minimization of eigen entropy and b) minimization
of dimensionality entropy objective functions for-
mulated in equations (7, 12) respectively. These two
objective functions are reviewed in the literature (De-
mantk
´
e et al., 2011; Weinmann et al., 2015), where
achieving lower entropy implies the choice of more
optimal parameters. The entropy minimization evalu-
ations demonstrate the pipeline’s efficiency of its op-
timization characteristics and ability to generalize for
learning over only a small subset of points in the PCD.
Despite the fact that neighborhood estimation is
explicitly correlated to the normal estimation task for
a PCD, we, additionally, use our pipeline to inves-
tigate the impact of optimal neighborhood selection
on such a task . In our work, the normal vector at
a point in the PCD is approximated by the eigenvec-
tor of the smallest eigenvalue of the estimated neigh-
borhood at the point. Besides, the unoriented angle
similarity and the proportion of good points (PGP) as
introduced in (Wang and Prisacariu, 2020) are used
to assess the quality of the normal estimation. As
seen in equation (16), the unoriented angle similarity
measures the similarity between the estimated normal
vector ˆn and a ground truth normal vector n, which
is the normal vector at the triangular face from which
the point is sampled. The unoriented angle similarity
calculates the similarity by first computing the angle β
between ˆn and n as in equation (15), and then normal-
izing the angle β by the maximum angular difference
β
max
, which in this case is set to 90
. The percent-
age of good points (PGP) is another evaluation metric
for normal estimation tasks, where PGP
α
denotes the
percentage of points having a normal estimation er-
ror of less than α, as seen in equation (17). Crucially,
we assess the proposed pipeline over normal estima-
tion task only since it has the aforementioned metrics
of evaluation, whereas keypoint detection and fea-
ture description have no quantitative evaluations and
are instead evaluated qualitatively which is inconclu-
sive. Furthermore, we compare our pipeline against
the classical Principal Components Analysis (PCA),
as discussed in (Demantk
´
e et al., 2011), since it com-
putes normal vectors in an unsupervised manner and
our pipeline serves as a pre-processing step for choos-
ing meaningful parameters to be used by PCA.
β = arccos
ˆn · n
k
ˆn
kk
n
k
(15)
similarity (β) = 1
β
β
max
(16)
PGP
α
=
1
N
aP
1
β
a
<α
(17)
4.3 Training Setting
The implementation of the proposed pipeline is real-
ized in Python 3.9 and PyTorch 1.10.0 without uti-
lizing GPUs or parallel threading. Even though a
GPU-enabled implementation is available and a par-
allel threading version of the pipeline, in which dif-
ferent keypoints can be deployed on multiple threads,
can be adapted in a straightforward manner for faster
execution. The lightweight PointNet blocks have 16
output features each, where the network weights are
initialized using Xavier uniform initialization and the
biases are set to 0.01. The hyperparameter σ of the
mutation function, as described in equation (14), is
set to 0.3 and decay linearly with 0.05 at each genera-
tion. Given each keypoint, the modified deep Neuro-
Evolution runs for 5 generations with a population
size of 50 individuals and keeps the top 20 perform-
ing individuals at each generation to act as parents of
the subsequent generation.
4.4 Results
Table 1 illustrates the eigen and dimensionality en-
tropy produced by our pipeline in different neighbor-
hood types, compared to the classical baseline. This
table is dedicated to the results of the training and test-
ing points. Training points are points sampled with
FPS and passed in an unsupervised training manner
to Deep Neuro-evolution. A total of 1028 points are
sampled from each mesh, however, only 51 points are
used for training. The findings indicate that indepen-
dent of the neighborhood type, our proposed pipeline
attains lower (better) entropy values compared to the
baseline. The test points are the 95% remaining points
that were not included in the pipeline training. Re-
sults show that the pipeline has a good generaliza-
tion performance over the unseen points. These re-
sults support our hypothesis that when directed by the
novel fitness function provided in equation (13), an
intelligent sampling approach (FPS) and Deep Neuro-
Evolution as a learning algorithm an effective and
novel pipeline for neighborhood estimation is pro-
posed.
Table 2 illustrates normal estimation results under
unoriented angle similarity and PGP metrics. PGP
is investigated at three distinct α levels in order to
offer a more comprehensive view of our approach’s
performance. The table draws a comparison between
our proposed pipeline trained with different types of
neighborhoods and optimizing either dimensionality
or eigenvalues entropy objective functions and the
baseline. Moreover, it provides results over training
points gathered with a sample percentage as discussed
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
590
Table 1: Entropy results on training/testing points between classical baseline vs. adaptable neighborhoods, estimated using
our pipeline for both spherical and K nearest neighborhoods. The lower values indicate better performance.
Neighborhood Type Eigen Entropy (train / test) Dimensionality Entropy (train / test)
Classic 0.8279 / 0.8279 0.8443 / 0.8436
Spherical Neighborhood 0.7365 / 0.7231 0.645 / 0.6456
K nearest neighbors 0.7185 / 0.7127 0.6431 / 0.6295
Table 2: Normal estimation results on training points between classical baseline vs. adaptable neighborhoods estimated using
our pipeline for both spherical and K nearest neighborhoods. The higher values indicate better performance.
Neighborhood Type Objective Function
Angular
Similarity
PGP 5 PGP 10 PGP 30
Classic N/A 0.6769 0.2029 0.3380 0.6140
Spherical Neighborhood
Dimension Entropy 0.6849 0.2542 0.3839 0.6211
Eigen Entropy 0.6953 0.2771 0.4093 0.6421
K nearest neighbors
Dimension Entropy 0.6856 0.2563 0.3865 0.6213
Eigen Entropy 0.6856 0.2563 0.3865 0.6213
Table 3: Normal estimation results on test points between classical baseline vs. adaptable neighborhoods estimated using our
pipeline for both spherical and K nearest neighborhoods. The higher values indicate better performance.
Neighborhood Type Objective Function
Angular
Similarity
PGP 5 PGP 10 PGP 30
Classic N/A 0.6778 0.2038 0.3384 0.6169
Spherical Neighborhood
Dimension Entropy 0.6880 0.2550 0.3849 0.6258
Eigen Entropy 0.6981 0.2804 0.4127 0.6463
K nearest neighbors
Dimension Entropy 0.6773 0.2070 0.3424 0.6167
Eigen Entropy 0.6773 0.2070 0.3424 0.6167
earlier. The results reveal that optimal neighborhood
selection has a positive influence on the normal es-
timation results. The proposed pipeline outperforms
the classical approach by a significant margin, with
the spherical neighborhood trained by eigen entropy
as the objective function being the best-performing
setting. When compared to the baseline, this set-
ting provides a 2% improvement in unoriented angu-
lar similarity and a 4% improvement in PGP on aver-
age.
The results are extended in Table 3 to include un-
seen test points. The results confirm that the spherical
neighborhood has the highest capacity to generalize
across unseen data. Hence, we can conclude based on
both Tables 2, 3 that the best fitting choice for nor-
mal estimation is a symmetric (spherical) neighbor-
hood with an equivalent dispersion in all directions
selected to minimize eigen entropy.
5 FUTURE WORK
For future work, it should be investigated whether the
learned neighborhood estimation can generalize not
only on points in the same PCD but also on points
from different PCDs that have similar structure or be-
long to the same class. The proposed approach may
additionally be validated on other datasets similar to
ModelNet such as PCPNet (Guerrero et al., 2018) and
is planned to be addressed in future works. Addition-
ally, the pipeline should be extended and validated on
large-scale PCDs with millions of points. Our hypoth-
esis is that employing our pipeline will have a positive
effect, as empirical results provided in LPDNet (Liu
et al., 2019) have shown that an optimal neighborhood
selection has a positive effect on the network perfor-
mance, yet the grid search used in this work has a
large memory and computation overhead. Different
sampling techniques should be investigated on a large
scale PCDs. One good candidate is using spatial clus-
tering over a PCD to create coarse regions of interest,
followed by FPS over the clustered areas. The ini-
tial clustering stage has the advantage of providing
an approximate count of the number of objects in the
scene, which helps FPS to sample points that summa-
rize the structure of the different objects. Finally, the
design of neighborhood selection heuristics and ob-
jective functions is still a work in progress, which can
be tailored to reduce errors in neighborhood attributes
like the normal vector.
6 CONCLUSION
In this work, we proposed a novel and efficient
pipeline for neighborhood estimation given PCDs,
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
591
which is a major challenge for both machine learning
and deep learning methods. The proposed pipeline
uses a modification of Deep Neuro-Evolution as a
learning module and FPS for intelligent sampling.
Besides, a novel fitness function is proposed to evalu-
ate the quality of each individual to reveal the elite
solution by applying random mutations of the net-
work parameters for a pre-determined number of gen-
erations. To further improve pipeline efficiency, we
propose to use only 5% of sampled points and uti-
lize them for the learning phase, while the remain-
ing 95% of points’ neighborhood is used to estimate
the pipeline generalization performance. In compar-
ison to the baseline, the pipeline was able to reduce
entropy values regardless of the neighborhood type.
Furthermore, the pipeline has a positive impact on
the normal estimation problem, with spherical neigh-
borhoods that optimize eigenvalues entropy deliver-
ing the best results.
REFERENCES
Bello, S. A., Yu, S., Wang, C., Adam, J. M., and Li, J.
(2020). Deep learning on 3d point clouds. Remote
Sensing, 12(11):1729.
Ben-Shabat, Y. and Gould, S. (2020). Deepfit: 3d surface
fitting via neural network weighted least squares. In
European conference on computer vision, pages 20–
34. Springer.
Ben-Shabat, Y., Lindenbaum, M., and Fischer, A. (2019).
Nesti-net: Normal estimation for unstructured 3d
point clouds using convolutional neural networks. In
Proceedings of the IEEE/CVF Conference on Com-
puter Vision and Pattern Recognition, pages 10112–
10120.
Chai, J., Zeng, H., Li, A., and Ngai, E. W. (2021). Deep
learning in computer vision: A critical review of
emerging techniques and application scenarios. Ma-
chine Learning with Applications, 6:100134.
Cirujeda, P., Mateo, X., Dicente, Y., and Binefa, X. (2014).
Mcov: a covariance descriptor for fusion of texture
and shape features in 3d point clouds. In 2014 2nd In-
ternational Conference on 3D Vision, volume 1, pages
551–558. IEEE.
Demantk
´
e, J., Mallet, C., David, N., and Vallet, B. (2011).
Dimensionality based scale selection in 3d lidar point
clouds. In Laserscanning.
Eldar, Y., Lindenbaum, M., Porat, M., and Zeevi, Y. Y.
(1994). The farthest point strategy for progressive im-
age sampling. Proceedings of the 12th IAPR Inter-
national Conference on Pattern Recognition, Vol. 2 -
Conference B: Computer Vision & Image Processing.
(Cat. No.94CH3440-5), pages 93–97 vol.3.
Fiolka, T., St
¨
uckler, J., Klein, D. A., Schulz, D., and
Behnke, S. (2012). Sure: Surface entropy for dis-
tinctive 3d features. In International Conference on
Spatial Cognition, pages 74–93. Springer.
Flint, A., Dick, A., and Van Den Hengel, A. (2007). Thrift:
Local 3d structure recognition. In 9th Biennial Con-
ference of the Australian Pattern Recognition Society
on Digital Image Computing Techniques and Applica-
tions (DICTA 2007), pages 182–188. IEEE.
Glorot, X. and Bengio, Y. (2010). Understanding the diffi-
culty of training deep feedforward neural networks. In
Proceedings of the thirteenth international conference
on artificial intelligence and statistics, pages 249–
256. JMLR Workshop and Conference Proceedings.
Guerrero, P., Kleiman, Y., Ovsjanikov, M., and Mitra, N. J.
(2018). Pcpnet learning local shape properties from
raw point clouds. In Computer Graphics Forum, vol-
ume 37, pages 75–85. Wiley Online Library.
Huang, J. and You, S. (2012). Point cloud matching based
on 3d self-similarity. In 2012 IEEE Computer Society
Conference on Computer Vision and Pattern Recogni-
tion Workshops, pages 41–48. IEEE.
Johnson, A. and Hebert, M. (1998). Surface matching for
object recognition in complex 3-d scenes. to appear in.
Image and Vision Computing.
Lazaros, N., Sirakoulis, G. C., and Gasteratos, A. (2008).
Review of stereo vision algorithms: From software to
hardware. International Journal of Optomechatronics,
2:435 – 462.
Liu, Z., Zhou, S., Suo, C., Yin, P., Chen, W., Wang, H., Li,
H., and Liu, Y.-H. (2019). Lpd-net: 3d point cloud
learning for large-scale place recognition and envi-
ronment analysis. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages
2831–2840.
Marton, Z.-C., Pangercic, D., Blodow, N., Kleinehellefort,
J., and Beetz, M. (2010). General 3d modelling of
novel objects from a single view. In 2010 ieee/rsj in-
ternational conference on intelligent robots and sys-
tems, pages 3700–3705. IEEE.
Ming, Y., Meng, X., Fan, C., and Yu, H. (2021). Deep
learning for monocular depth estimation: A review.
Neurocomputing, 438:14–33.
Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2017a). Point-
net: Deep learning on point sets for 3d classification
and segmentation. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 652–660.
Qi, C. R., Yi, L., Su, H., and Guibas, L. J. (2017b). Point-
net++: Deep hierarchical feature learning on point sets
in a metric space. Advances in neural information pro-
cessing systems, 30.
Rabbani, T., Van Den Heuvel, F., and Vosselmann, G.
(2006). Segmentation of point clouds using smooth-
ness constraint. International archives of photogram-
metry, remote sensing and spatial information sci-
ences, 36(5):248–253.
Royo, S. and Ballesta-Garcia, M. (2019). An overview of
lidar imaging systems for autonomous vehicles. Ap-
plied Sciences, 9:4093.
Rusu, R. B., Blodow, N., Marton, Z. C., and Beetz, M.
(2008a). Aligning point cloud views using persistent
feature histograms. In 2008 IEEE/RSJ international
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
592
conference on intelligent robots and systems, pages
3384–3391. IEEE.
Rusu, R. B., Marton, Z. C., Blodow, N., and Beetz, M.
(2008b). Persistent point feature histograms for 3d
point clouds. In Proc 10th Int Conf Intel Autonomous
Syst (IAS-10), Baden-Baden, Germany, pages 119–
128.
Salimans, T., Ho, J., Chen, X., Sidor, S., and Sutskever,
I. (2017). Evolution strategies as a scalable al-
ternative to reinforcement learning. arXiv preprint
arXiv:1703.03864.
Siritanawan, P., Prasanjith, M. D., and Wang, D. (2017).
3d feature points detection on sparse and non-uniform
pointcloud for slam. In 2017 18th International Con-
ference on Advanced Robotics (ICAR), pages 112–
117. IEEE.
Such, F. P., Madhavan, V., Conti, E., Lehman, J., Stanley,
K. O., and Clune, J. (2017). Deep neuroevolution: Ge-
netic algorithms are a competitive alternative for train-
ing deep neural networks for reinforcement learning.
arXiv preprint arXiv:1712.06567.
Suo, C., Liu, Z., Mo, L., and Liu, Y. (2020). Lpd-ae: la-
tent space representation of large-scale 3d point cloud.
IEEE Access, 8:108402–108417.
Toshev, A., Mordohai, P., and Taskar, B. (2010). Detecting
and parsing architecture at city scale from range data.
In 2010 IEEE Computer Society Conference on Com-
puter Vision and Pattern Recognition, pages 398–405.
IEEE.
Triebel, R., Kersting, K., and Burgard, W. (2006). Robust
3d scan point classification using associative markov
networks. In Proceedings 2006 IEEE International
Conference on Robotics and Automation, 2006. ICRA
2006., pages 2603–2608. IEEE.
Vandapel, N., Huber, D. F., Kapuria, A., and Hebert, M.
(2004). Natural terrain classification using 3-d ladar
data. In IEEE International Conference on Robotics
and Automation, 2004. Proceedings. ICRA’04. 2004,
volume 5, pages 5117–5122. IEEE.
Varga, D., Szalai-Gindl, J. M., Formanek, B., Vaderna, P.,
Dobos, L., and Laki, S. (2021). Template matching
for 3d objects in large point clouds using dbms. IEEE
Access, 9:76894–76907.
Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M.,
and Solomon, J. M. (2019). Dynamic graph cnn
for learning on point clouds. Acm Transactions On
Graphics (tog), 38(5):1–12.
Wang, Z. and Prisacariu, V. A. (2020). Neighbourhood-
insensitive point cloud normal estimation network.
arXiv preprint arXiv:2008.09965.
Weinmann, M., Jutzi, B., and Mallet, C. (2014). Seman-
tic 3d scene interpretation: A framework combining
optimal neighborhood size selection with relevant fea-
tures. ISPRS Annals of the Photogrammetry, Remote
Sensing and Spatial Information Sciences, 2(3):181.
Weinmann, M., Schmidt, A., Mallet, C., Hinz, S., Rotten-
steiner, F., and Jutzi, B. (2015). Contextual classifi-
cation of point cloud data by exploiting individual 3d
neigbourhoods. ISPRS Annals of the Photogramme-
try, Remote Sensing and Spatial Information Sciences
II-3 (2015), Nr. W4, 2(W4):271–278.
Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X.,
and Xiao, J. (2015). 3d shapenets: A deep representa-
tion for volumetric shapes. In Proceedings of the IEEE
conference on computer vision and pattern recogni-
tion, pages 1912–1920.
Yang, Y., Feng, C., Shen, Y., and Tian, D. (2018). Fold-
ingnet: Point cloud auto-encoder via deep grid defor-
mation. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 206–
215.
Zhao, C., Yang, J., Xiong, X., Zhu, A., Cao, Z., and Li,
X. (2022). Rotation invariant point cloud analysis:
Where local geometry meets global topology. Pattern
Recognition, 127:108626.
Zhong, Y. (2009). Intrinsic shape signatures: A shape de-
scriptor for 3d object recognition. In 2009 IEEE 12th
international conference on computer vision work-
shops, ICCV Workshops, pages 689–696. IEEE.
Zhou, J., Jin, W., Wang, M., Liu, X., Li, Z., and Liu, Z.
(2021a). Improvement of normal estimation for point-
clouds via simplifying surface fitting. arXiv preprint
arXiv:2104.10369.
Zhou, J., Jin, W., Wang, M., Liu, X., Li, Z., and Liu,
Z. (2021b). Improvement of normal estimation for
pointclouds via simplifying surface fitting. ArXiv,
abs/2104.10369.
Zhu, R., Liu, Y., Dong, Z., Wang, Y., Jiang, T., Wang, W.,
and Yang, B. (2021). Adafit: Rethinking learning-
based normal estimation on point clouds. In Pro-
ceedings of the IEEE/CVF International Conference
on Computer Vision, pages 6118–6127.
Point Cloud Neighborhood Estimation Method Using Deep Neuro-Evolution
593