Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification Lifa Zhu1 Changwei Lin1 Chen Zheng2 and Ninghua Yang1

2025-05-02 0 0 903.38KB 10 页 10玖币

侵权投诉

Point-Voxel Adaptive Feature Abstraction for

Robust Point Cloud Classiﬁcation

Lifa Zhu1, Changwei Lin1, Chen Zheng2, and Ninghua Yang1

1DeepGlint, Beijing, China

2China University of Mining and Technology, Beijing, China

zhulf0804@gmail.com

{changweilin,ninghuayang}@deepglint.com

zc824435664@163.com

Abstract. Great progress has been made in point cloud classiﬁcation

with learning-based methods. However, complex scene and sensor inaccu-

racy in real-world application make point cloud data suﬀer from corrup-

tions, such as occlusion, noise and outliers. In this work, we propose

Point-Voxel based Adaptive (PV-Ada) feature abstraction for robust

point cloud classiﬁcation under various corruptions. Speciﬁcally, the pro-

posed framework iteratively voxelize the point cloud and extract point-

voxel feature with shared local encoding and Transformer. Then, adap-

tive max-pooling is proposed to robustly aggregate the point cloud fea-

ture for classiﬁcation. Experiments on ModelNet-C dataset demonstrate

that PV-Ada outperforms the state-of-the-art methods. In particular, we

rank the 2nd place in ModelNet-C classiﬁcation track of PointCloud-C

Challenge 2022, with Overall Accuracy (OA) being 0.865. Code will be

available at https://github.com/zhulf0804/PV-Ada.

Keywords: point-voxel, adaptive feature abstraction, robust classiﬁca-

tion

1 Introduction

Point cloud classiﬁcation, as one of the important 3D tasks, has achieved remark-

able progress[15,16,22,8,23,3,28,11,12]. However, the state-of-the-art methods

are less robust in 3D point cloud recognition under corruptions or attack[17,21,20,14].

In this report, we mainly describe our method and experiments on ModelNet-C

point cloud classiﬁcation hosted in PointCloud-C Challenge 20223, which bench-

marks point cloud robustness analysis under corruptions[17].

ModelNet-C is the ﬁrst systematically-designed test-suite for point cloud clas-

siﬁcation under corruptions[17]. It extends the validation set of ModelNet40 with

the proposed atomic corruptions, including “Add Global”, “Add Local”, “Drop

Global”, “Drop Local”, “Rotate”, “Scale” and “Jitter”. Moreover, each atomic

corruption is performed with 5 severity levels. It’s noted that each real-world

corruption can be broken down into a combination of the atomic corruptions.

3https://pointcloud-c.github.io/competition.html

arXiv:2210.15514v2 [cs.CV] 30 Oct 2022

2 L. Zhu et al.

In terms of the corruptions in ModelNet-C data, we propose Point-Voxel

Adaptive (PV-Ada) feature abstraction for robust point cloud classiﬁcation. PV-

Ada is proposed based on the following two observations. First, the rough con-

tour may beneﬁt point cloud classiﬁcation under corruptions, instead of paying

attention to detailed structures. It inspires us to involve both voxel-level and

point-level feature for point cloud classiﬁcation. Second, not all all points in

point cloud are equal for individual feature representation, such as plane points

and outlier points. So we propose to learn point weight in an end-to-end manner,

thus conducting adaptive pooling for robust feature abstraction.

Experiments on ModelNet-C and ModelNet demonstrate the eﬀectiveness

of PV-Ada. Our method outperforms the state-of-the-art published models and

achieves 0.884 and 0.865 OA on the public and private ModelNet-C test set,

respectively.

2 Method

A point cloud is represented as a set X={xi∈R3}i=1,2,··· ,N , where Nis the

point number in X. For the convenience of following discussion, point cloud is

denoted by matrix X∈RN×3. Inputted point cloud X, PV-Ada outputs C

scores for all the Ccandidate categories. The overall architecture of PV-Ada is

depicted in Fig. 1.

2.1 Point-voxel encoder

Point-voxel as input and voxelization operation have been widely used in point

cloud detection, segmentation and classiﬁcation[18,19,10,24,7,9,22,2,1], which

shows the eﬀectiveness and eﬃciency of voxelization. Our presented point-voxel

encoder consists of three modules: local encoding, Transformer and pyramid

feature interaction.

Local encoding Local encoding groups knearest neighbors for each point (voxel),

then learns Ddimensional point (voxel) feature through convolution and non-

linear operations. Take X∈RN×3as an example, local encoding outputs its

local desciptor Fe

X∈RN×D.

Transformer We adopt PCT[3] to enhance the point feature for point cloud.

Take Xas an example, it ﬁrst further encode point feature Fe

Xwith MLP and

non-linear operations, generating Ddimensional features Fm

X∈RN×D. Then

four oﬀset-attention are introduced to obtain features (Foa1

X,Foa2

X,Foa3

X,Foa4

X)∈

RN×D. Each is generated through a vanilla attention and a oﬀset-based residual

structure: Foa0

X=Fm

Fvai

X= VanillaAttention(Foai−1

X), i = 1,2,3,4,

Foai

X=Foai−1

X+fi(Foai−1

X−Fvai

X),

(1)

where fi(·) is implemented using Conv1d, normalization and ReLU.

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：

Point-VoxelAdaptiveFeatureAbstractionforRobustPointCloudClassificationLifaZhu1,ChangweiLin1,ChenZheng2,andNinghuaYang11DeepGlint,Beijing,China2ChinaUniversityofMiningandTechnology,Beijing,Chinazhulf0804@gmail.com{changweilin,ninghuayang}@deepglint.comzc824435664@163.comAbstract.Greatprogresshasbeenm...

展开>> 收起<<

Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification Lifa Zhu1 Changwei Lin1 Chen Zheng2 and Ninghua Yang1.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！

Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification Lifa Zhu1 Changwei Lin1 Chen Zheng2 and Ninghua Yang1

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: