Centroid Distance Keypoint Detector for Colored Point Clouds Hanzhe Teng Dimitrios Chatziparaschis Xinyue Kan Amit K. Roy-Chowdhury Konstantinos Karydis University of California Riverside

2025-04-27 0 0 6.25MB 10 页 10玖币

侵权投诉

Centroid Distance Keypoint Detector for Colored Point Clouds

Hanzhe Teng, Dimitrios Chatziparaschis, Xinyue Kan, Amit K. Roy-Chowdhury, Konstantinos Karydis

University of California, Riverside

United States of America

{hteng007, dchat013, xkan001, amitrc, karydis}@ucr.edu *

Abstract

Keypoint detection serves as the basis for many com-

puter vision and robotics applications. Despite the fact that

colored point clouds can be readily obtained, most existing

keypoint detectors extract only geometry-salient keypoints,

which can impede the overall performance of systems that

intend to (or have the potential to) leverage color informa-

tion. To promote advances in such systems, we propose an

efﬁcient multi-modal keypoint detector that can extract both

geometry-salient and color-salient keypoints in colored

point clouds. The proposed CEntroid Distance (CED) key-

point detector comprises an intuitive and effective saliency

measure, the centroid distance, that can be used in both 3D

space and color space, and a multi-modal non-maximum

suppression algorithm that can select keypoints with high

saliency in two or more modalities. The proposed saliency

measure leverages directly the distribution of points in a

local neighborhood and does not require normal estima-

tion or eigenvalue decomposition. We evaluate the pro-

posed method in terms of repeatability and computational

efﬁciency (i.e. running time) against state-of-the-art key-

point detectors on both synthetic and real-world datasets.

Results demonstrate that our proposed CED keypoint de-

tector requires minimal computational time while attaining

high repeatability. To showcase one of the potential ap-

plications of the proposed method, we further investigate

the task of colored point cloud registration. Results sug-

gest that our proposed CED detector outperforms state-of-

the-art handcrafted and learning-based keypoint detectors

in the evaluated scenes. The C++ implementation of the

proposed method is made publicly available at https://

github.com/UCR-Robotics/CED_Detector.

*We gratefully acknowledge the support of NSF grants #IIS-1724341

and #IIS-1901379, ONR grant #N00014-18-1-2252, and USDA/NIFA

#2021-67022-33453. Any opinions, ﬁndings, and conclusions or recom-

mendations expressed in this material are those of the authors and do not

necessarily reﬂect the views of the funding agencies.

1. Introduction

Keypoint detection serves as the basis across computer

vision and robotics applications such as 3D reconstruc-

tion [1,11], localization and mapping [8,12], and navigation

on point clouds [9, 17]. In such applications, colored point

clouds can be obtained from RGB-D cameras, where color

and depth images are aligned, or camera-LiDAR systems,

where they are collected after extrinsic calibration; this pro-

cess is often referred to as point cloud colorization or pho-

tometric reconstruction (e.g., [21]). However, most existing

keypoint detectors only consider geometric information and

fail to extract color-rich keypoints, which impedes the over-

all performance of systems that intend to (or have the poten-

tial to) leverage color information. Hence, there is need for

efﬁcient keypoint detectors that can extract both geometric

and color information in the point cloud.

Keypoint detection aims to extract a subset of points

from the point cloud so that they can best represent the

original data in a compact form. Some successful key-

point detectors for 2D images, such as SIFT [19] and Har-

ris [10], have been extended to 3D space following the orig-

inal design ideas by the Point Cloud Library (PCL) com-

munity [28]. However, the data structures used in 2D im-

ages and 3D point clouds are fundamentally different. This

may limit deployment of such methods and has led to stud-

ies that focus on intrinsic properties of point clouds. Re-

cent advances in deep learning has helped introduce sev-

eral learning-based keypoint detectors and feature descrip-

tors [6, 18, 37, 39]. In spite of their strong performance

within their training domains, learned detectors and features

may not transfer over in new scenes that are different from

those used in training. For example, it may be challenging

for a system trained with data collected in indoor environ-

ments to operate in diverse outdoor scenes [2].

In contrast, methods that leverage the inherent properties

of point clouds may help overcome this difﬁculty. Several

existing methods focusing on geometric properties of point

clouds, such as NARF [29] and ISS [40], require eigenvalue

decomposition and/or normal estimation. These operations

are computationally expensive, especially when performing

arXiv:2210.01298v2 [cs.CV] 15 Jun 2023

keypoint detection at a large scale. In a distinct line of re-

search, it has been shown that incorporating color modality

can improve accuracy for applications such as point cloud

registration [23].

The main hypothesis underlying this work is that in-

corporating color modality (in addition to a geometric

modality) can help improve the overall performance, as the

amount of information passed on to the following compo-

nents in the system has increased. Despite existing descrip-

tors that can incorporate color information (e.g., [32]), to

the best of the authors’ knowledge there currently exists no

effective keypoint detector that can extract color-rich key-

points to feed to the descriptor. For instance, geometric-

based keypoint detectors can fail to extract keypoints on a

ﬂat surface with color texture. While some methods (e.g.,

SIFT-3D) can extract color-rich keypoints, they do so at ex-

pense of losing geometric information (i.e. they can only

respond to one modality). This limitation can be linked to

lack of an effective non-maximum suppression algorithm to

combine the two modalities; this is one of the key contribu-

tions of this work.

To this end, we propose an efﬁcient multi-modal key-

point detector, named CEntroid Distance (CED) keypoint

detector, that utilizes both geometric and photometric in-

formation. The proposed CED detector comprises an intu-

itive and effective saliency measure, the centroid distance,

that can be used in both 3D space and color space, and

a multi-modal non-maximum suppression algorithm that

can select keypoints with high saliency in two or more

modalities. The proposed saliency measure leverages di-

rectly the distribution of points in a local neighborhood and

does not require normal estimation or eigenvalue decom-

position. The proposed CED detector is evaluated in terms

of repeatability and computational efﬁciency (running time)

against state-of-the-art keypoint detectors on both synthetic

and real-world datasets. Results demonstrate that our pro-

posed CED keypoint detector requires minimal computa-

tional time while attaining high repeatability. In addition,

to showcase one of the potential applications of the pro-

posed method, we further investigate the task of colored

point cloud registration. Results show that our CED detec-

tor outperforms state-of-the-art crafted and learning-based

keypoint detectors in the evaluated scenes.

Contributions. The paper’s contributions are fourfold:

• We propose an efﬁcient multi-modal keypoint detector

that can extract both geometry-salient and color-salient

keypoints in a colored point cloud, with the potential to

be extended and applied to point clouds with multiple

modalities (e.g., colored by multi-spectrum images).

• We propose to use an intuitive and effective measure

for keypoint saliency, the distance to centroid, which

can leverage directly the distribution of points and does

not require normal estimation or eigenvalue decompo-

sition.

• We develop a multi-modal non-maximum suppression

algorithm that can select keypoints with high saliency

in two or more modalities.

• We demonstrate through experiments in four datasets

that the proposed keypoint detector can outperform the

state-of-the-art handcrafted and learning-based key-

point detectors.

2. Related Work

3D keypoint detectors can be categorized as those ex-

tending designs originally developed for 2D images [10,19],

and those native to 3D point clouds [4, 20, 31, 40] and 3D

meshes [3, 34, 38]. Following the design in 2D images,

Harris family [10] computes covariance matrices of surface

normal or intensity gradient in 3D space, and in 3D and

color space (herein referred to as 6D space). SIFT [19]

applies the difference of Gaussian operator in scale-space

to ﬁnd keypoints with local maximal response. However,

for 3D point clouds, the amount and position of points

within the spherical region are uncertain, making it hard

to obtain gradients. In 3D space, Normal Aligned Ra-

dial Feature (NARF) [29] measures saliency from surface

normal and distance changes between neighboring points.

Intrinsic Shape Signature (ISS) [40] and KeyPoint Qual-

ity (KPQ) [20] perform eigenvalue decomposition of the

scatter matrix of neighbor points and threshold on the ra-

tio between eigenvalues. Heat Kernel Signature (HKS) [31]

and Laplace-Beltrami Scale-space (LBSS) [34] measure the

saliency from the response to the Laplace-Beltrami opera-

tor in the neighborhood. Local Surface Patches (LSP) [4]

leverages local principal curvatures to construct the Shape

Index (SI) [7] as the measure of saliency. As in SIFT,

MeshDoG [38] and Salient Points (SP) [3] apply the

difference-of-Gaussian operator to construct the scale space

for saliency measure. We refer readers to the comprehen-

sive evaluation in [33] for more details.

In summary, the existing methods often apply an oper-

ator to obtain point normal, curvature and gradient in the

local region, and threshold on either a combination of the

obtained measures or the eigenvalues of the covariance ma-

trices. On the contrary, our proposed method leverages di-

rectly the point distribution and statistics in 3D space and

color space, without the need of normal estimation or eigen-

value decomposition.

Learning-based approaches, such as USIP [18] and

3DFeat-Net [37], have also been studied. 3DFeat-Net [37]

learns a 3D feature detector and descriptor for point cloud

matching using weak supervision, whereas USIP [18] trains

a feature proposal network with probabilistic Chamfer loss

in an unsupervised manner.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CentroidDistanceKeypointDetectorforColoredPointCloudsHanzheTeng,DimitriosChatziparaschis,XinyueKan,AmitK.Roy-Chowdhury,KonstantinosKarydisUniversityofCalifornia,RiversideUnitedStatesofAmerica{hteng007,dchat013,xkan001,amitrc,karydis}@ucr.edu*AbstractKeypointdetectionservesasthebasisformanycom-puterv...

展开>> 收起<<

Centroid Distance Keypoint Detector for Colored Point Clouds Hanzhe Teng Dimitrios Chatziparaschis Xinyue Kan Amit K. Roy-Chowdhury Konstantinos Karydis University of California Riverside.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Centroid Distance Keypoint Detector for Colored Point Clouds Hanzhe Teng Dimitrios Chatziparaschis Xinyue Kan Amit K. Roy-Chowdhury Konstantinos Karydis University of California Riverside

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: