Large-scale 3D Point Cloud Classification Based On Feature
Description Matrix By CNN
*
Lei Wang
†
School of Computer, Northwestern
Polytechnical University
1
,
East China University of Technology
2
wlei598@163.com
Weiliang Meng
LIAMA - NLPR, CAS Institute of
Automation
weiliang.meng@ia.ac.cn
Runping Xi
‡
School of Computer, Northwestern
Polytechnical University
xrp@163.com
Yanning Zhang
S
School of Computer, Northwestern
Polytechnical University
ynzhang@nwpu.edu.cn
Ling Lu
East China University of Technology
luling2006@163.com
Xiaopeng Zhang
LIAMA - NLPR, CAS Institute of
Automation
xiaopeng.zhang@ia.ac.cn
ABSTRACT
Large-scale 3D Point cloud classification is a basic topic for various
applications. Traditional geometries features are usually indepen-
dent of each other and difficult to adapt to a fixed classification
model. With the rise of the neural network, deep learning is consid-
ered in 3D point cloud application. 3D points are difficult to feed
the neural network directly based on deep learning, as they can-
not be arranged in a fixed order as image pixels. In this paper, we
combine traditional feature-based methods with the Convolution-
al neural network(CNN) to finish the classification task. The core
idea is to construct a feasible structure called Feature Description
Matrix(FDM) which encapsulates the local feature of the point to
feed CNN for training and testing. By extracting geometry features
and designed Feature Description Vectors(FDV) for FDM, a simple
mechanism for point cloud classification is given, and experiments
validate the effectiveness of our method, with higher classification
accuracy compared to state-of-art works.
KEYWORDS
point cloud, feature extraction, deep learning, feature description
matrix
ACM Reference Format:
Lei Wang, Weiliang Meng, Runping Xi, Yanning Zhang, Ling Lu, and Xi-
aopeng Zhang. 2018. Large-scale 3D Point Cloud Classification Based On
Feature Description Matrix By CNN. In CASA 2018: 31st International
*
This work is supported in part by National Natural Science Foundation of China with
Nos. 61561003, 61571439, 61572405, 61761003, and in part by the Open Projects
Program of National Laboratory of Pattern Recognition with No.201600038 and Project
6140001010207.
†
Corresponding Author
‡
Corresponding Author
S
Corresponding Author
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
CASA 2018, May 21–23, 2018, Beijing, China
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-6376-1/18/05. . . $15.00
https://doi.org/10.1145/3205326.3205355
Conference on Computer Animation and Social Agents, May 21–23, 2018,
Beijing, China. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/
3205326.3205355
1 INTRODUCTION
Automatic analysis for 3D point cloud is more important in lots of
applications such as remote sensing, scene reconstruction, as the
obtaining of the 3D point cloud from real scene becomes cheap, con-
venient and fast. Currently, the semantic information is difficult to be
inferred from the point cloud directly, as the points are independent
of each other and no connections information can be employed. In
order to determine the semantic information for each point, we need
to classify the point cloud into different classes.
Many point cloud classification methods have been proposed
for different purposes. Usually, the features of each point must
be extracted firstly based on their local neighborhood, which is
related to the properties of geometry. The geometric properties of
natural surfaces may span over a wide range of scales (from
𝑐𝑚
to
𝑘𝑚
), and lots of works have been done on the natural scenes
understanding such as dune fields, 3D stratigraphic reconstruction
and outcrop analysis( [
19
]), grain size distribution in rivers ( [
15
]),
dune fields ( [
25
]), vegetation hydraulic roughness( [
1
]), channel bed
dynamics ( [
24
]) and in situ monitoring of cliff erosion and rockfall
characteristics ( [
20
]), or on the other scenes([
4
], [
14
], and [
10
]),
while different benchmarks are also be given( [32] and [13]).
Deep learning receives great interest in recent years because of
its excellent performance, especially on image recognition and un-
derstanding. The representative works for deep learning on 3D data
are volumetric CNN [38], 3DCNN-DQN-RNN [21], pointNet [28],
pointNet++ [
29
], O-CNN [
34
], and PCPNet [
11
] etc.. Although
deep learning methods can capture the features from the input im-
plicitly and generate the corresponding output labels after training,
the learning process cannot always work for every case. If some
features can be extracted explicitly for deep learning, a better classi-
fication result may be obtained. Based on this point, we propose a
new point cloud classification method which combines traditional
feature-based method with CNN. We first extracted a series of fea-
tures for each point based on their neighborhood, then construct a
Neighborhood Feature Matrix(FDM) to feed CNN in order to detect
the connections between the features and the corresponding labels in
turn to obtain invincible classification results. The core idea lies in