SIPF: SCALE INVARIANT POINT FEATURE FOR 3D POINT CLOUDS
Baowei Lin
1
, Fangda Zhao
2
, Toru Tamaki
2
, Fasheng Wang
1
,LeXiao
3
1
Dalian Neusoft University of Information, China
2
Hiroshima University, Japan
3
Dalian University of Technology, China
ABSTRACT
In this paper, we propose a method for detecting Scale-
Invariant Point Feature(SIPF) including 3D keypoints Detector
and feature descriptor. To detect SIPF, we first estimate a
keyscale for point cloud, and calculate the covariance matrix
of each 3D point. Keypoints are the saliency points who have
a fast change speed along with all principal directions. Then
the descriptors are encoded based on the shape of a border or
silhouette of an object to be detected or recognized. Experi-
mental results with the Stanford datasets demonstrate that the
proposed method can be effectively used for 3D point clouds
expression.
Index Terms— SIPF, scale invariant, feature, point cloud
1. INTRODUCTION
In this paper, we propose a novel method of scale-invariant
feature descriptor for 3D point clouds. Feature issue is very
important task in computer vision. A lot of 2D local fea-
ture methods have been imposed for image understanding,
detecting or matching targets by using the information of
images such as color, texture or differential pixel value.
Among them, SIFT [2] should be one of the most widely
used feature detector and descriptor because of the character-
istic of scale invariant.
Inspired by 2D local descriptors, recently researchers
have started to work with 3D local descriptors invariant to
rotation and translation of 3D point clouds. Based on the fast-
developed 3D acquisition techniques, point clouds become
an important representation of 3D scenes. Different from
2D image and mesh model, point cloud is a more natural
way for 3D representation without any assumption of con-
nectivity information or underlying topology. However,if
different acquisition devices or methods are used, most of the
reconstructed point clouds of the same scene have different
scale sizes. Even though various of 3D local descriptors have
been proposed, most of the feature descriptors are not scale
invariant, which is a critical issue for application like the
registration problems of the point clouds with different sizes.
Researchers have proposed several methods for describing
3D points. They can be mainly divided into two classes as
follows.
Non-Scale-Adaptive: One of the most popular 3D descriptor
is SPIN image [4, 5]. It is a 2D representation of the 3D
surface which surrounds the keypoint in a certain search
radius. If the neighbor size of search area can be established,
keypoint oriented descriptor can express local feature effi-
ciently. Point Feature Histograms (PFH) [6] and Fast Point
Feature Histograms (FPFH) [7] are good 3D feature descrip-
tors based on histograms. They represent the relationships
between neighbors of 3D keypoints and make them a good
bin expressed accumulation which is invariant to 3D point
cloud density. Steder et al. [3] have proposed a depth change
based 3D descriptor NARF. It extracts keypoints and descrip-
tors by using high changing spread along depth direction in
range image. Methods above are good for 3D local feature
description; however, they are not scale invariant, not even
used for a 3D point cloud.
Scale-Adaptive:There are some 3D scale invariant fea-
tures of extensions of 2D methods, such as 3D SIFT [8], 3D
SURF [9] and nD SIFT [10] which describe features of vol-
umes or n-dimensional data. However, these methods take
advantage of the mesh information of 3D scene. They are not
useful in a 3D point cloud.
Keyscale [18] is a good way to register point clouds with
different scale. This method estimates a keyscale to repre-
sent the scale size of target point cloud. By resizing the point
cloud with this scale, two point clouds share the same global
scale. However, scale information is not enough for encoding
descriptors.
Instead of directly extending those 2D features, some
researchers have proposed to combine 2D features with 3D
point clouds [17]. It works when 2D images are available,
but not applicable to characterize the descriptors of 3D point
clouds themselves.
Some texture information analysis based descriptors have
been proposed such as SHOT (signatures of histograms) [11]
and VIP (viewpoint invariant patches) [12]. The additional
information make the local feature more descriptive. These
methods are not for our non-information 3D point clouds.
Some scale-fixed descriptors can be used as scale invariant
detectors such as LBSS (Laplace Beltrami Scale Space) [13],
MeshDoG [14], Keypoint Quality Adaptive Scale [16] and
978-1-4799-8339-1/15/$31.00 ©2015 IEEE