MOTION ESTIMATION AND FILTERED PREDICTION FOR DYNAMIC POINT CLOUD ATTRIBUTE COMPRESSION Haoran Hong Eduardo Pavez Antonio Ortega Ryosuke Watanabey Keisuke Nonakay

2025-05-02 0 0 1022.86KB 5 页 10玖币

侵权投诉

MOTION ESTIMATION AND FILTERED PREDICTION FOR DYNAMIC POINT CLOUD

ATTRIBUTE COMPRESSION

Haoran Hong?, Eduardo Pavez?, Antonio Ortega?, Ryosuke Watanabe?,†, Keisuke Nonaka†

?University of Southern California, Los Angeles CA. 90089 USA

†KDDI Research, Inc., Japan

ABSTRACT

In point cloud compression, exploiting temporal redundancy for

inter predictive coding is challenging because of the irregular geom-

etry. This paper proposes an efﬁcient block-based inter-coding

scheme for color attribute compression. The scheme includes

integer-precision motion estimation and an adaptive graph based

in-loop ﬁltering scheme for improved attribute prediction. The pro-

posed block-based motion estimation scheme consists of an initial

motion search that exploits geometric and color attributes, followed

by a motion reﬁnement that only minimizes color prediction error.

To further improve color prediction, we propose a vertex-domain

low-pass graph ﬁltering scheme that can adaptively remove noise

from predictors computed from motion estimation with different

accuracy. Our experiments demonstrate signiﬁcant coding gain over

state-of-the-art coding methods.

Index Terms—inter prediction, geometry-based point cloud

coding, motion estimation, low-pass ﬁlter, graph ﬁlter.

1. INTRODUCTION

In current video coding systems, motion estimation (ME) and mo-

tion compensated prediction (MCP) are two essential techniques to

exploit temporal redundancy between frames, so that inter-predictive

coding techniques can be applied. In video-based point cloud com-

pression (V-PCC) [1] a direct extension of these motion-based tech-

niques to dynamic point cloud compression (PCC) is straightfor-

ward. However, for geometry-based PCC systems, where a point

cloud (PC) is encoded directly in 3D space, ME becomes challeng-

ing. To see why, note that within a block in a video frame all pixel

positions contain color information. Thus, both the reference and

predictor blocks have the same number of pixels. In contrast, since

PCs represent the surface of objects within a volume. Each 3D block

in general has a different geometry and a different number of occu-

pied voxels. Due to this irregular spatial structure, ME for geometry-

based PC coding has to rely on PC registration (PCR) techniques [2],

which match point sets of different sizes and geometries.

While PCR techniques have been widely studied and play a crit-

ical role in many computer vision applications, they need to be mod-

iﬁed for ME and MCP, where the goal is to achieve better temporal

prediction of PC attributes such as color. ME can be viewed as the

estimation of a 3D transformation between two PCs, which seeks

to align each element in a set of non-overlapping units in the pre-

dicted PC (e.g., voxels, blocks or clusters) to a corresponding unit in

the reference PC. As an example, each block in the predicted PC is

This work was funded in part by KDDI Research, Inc. and by the Na-

tional Science Foundation (NSF CNS-1956190).

matched to a corresponding block in the reference and the transfor-

mation that best aligns the blocks is the estimated motion. Existing

geometry-based ME techniques are based on iterative closest point

(ICP) [3,4,5], graph matching [6] and block matching [7,8,9].

ICP-based methods are the most commonly used due to their

good performance and efﬁciency. They work particularly well when

the two PCs to be matched are nearly aligned to begin with, which

is usually the case for consecutive frames in dynamic PCs captured

at high frame rates (e.g., 30 frames per second). In contrast, other

PCR methods for ME are more complex and cannot be easily scaled

to large PCs. Thus, our proposed ME is based on ICP.

MCP for geometry-based inter predictive coding can be accom-

plished by simply replacing the voxels in a predicted block with vox-

els in its best matching reference block [3,7], with mode decision

determining whether intra coding should be used instead if a sufﬁ-

ciently good approximation is not given by the copied block. As an

alternative, in [4,10] the color of each voxel in the current frame

is predicted by the color of its corresponding voxel in the reference

frame and a residual signal is sent. Our proposed method uses blocks

instead of clusters as the basic unit for ME (as [3,7]) but generates

residues (as [6,11,4]). In addition, we propose a vertex-domain low

pass graph ﬁlter that leads to signiﬁcant improvements in prediction.

Our work addresses several limitations in recent work. First, ex-

isting block-based ME schemes [3,7,8,9,5] mimic video coding

approaches by matching blocks of same size in the predicted and

reference frames. Instead, as in [4] we directly optimize the align-

ment of a block in the current frame to a larger bounding box in the

reference frame. With this approach, the closest points matched in

the reference frame are not restricted to be within a block of same

size as the block in the current frame. Note that using blocks is more

efﬁcient than using clusters, as in [4], since no clustering operations

are needed, and blocks can be efﬁciently obtained using the octree

structure. Also, different from recent ME schemes [4,3,5] we use

a modiﬁed color-ICP [12] without rotation transformation (to avoid

sending rotation information as overhead), with a hybrid matching

metric that combines geometry and color, which leads to better inter

prediction performance.

Even though our proposed block-based ME is based on a crite-

rion that combines color and geometry (using color-ICP [12]) there

is no guarantee that it is optimal for predicting PC color attributes.

Thus, after the initial block matching, we propose two novel tech-

niques not considered in the geometric point cloud coding literature.

First, we use a reﬁned local search around the color-ICP based mo-

tion to generate a reﬁned motion that minimizes the color predic-

tion error. By minimizing residual energy we improve prediction,

while preserving geometric consistency because we only search lo-

cally around the initial motion found through color-ICP. Second, for

MCP, we propose a vertex-domain low-pass graph ﬁlter. Unlike pre-

arXiv:2210.08262v2 [cs.CV] 28 Oct 2022

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：

MOTIONESTIMATIONANDFILTEREDPREDICTIONFORDYNAMICPOINTCLOUDATTRIBUTECOMPRESSIONHaoranHong?,EduardoPavez?,AntonioOrtega?,RyosukeWatanabe?;y,KeisukeNonakay?UniversityofSouthernCalifornia,LosAngelesCA.90089USAyKDDIResearch,Inc.,JapanABSTRACTInpointcloudcompression,exploitingtemporalredundancyforinterpred...

展开>> 收起<<

MOTION ESTIMATION AND FILTERED PREDICTION FOR DYNAMIC POINT CLOUD ATTRIBUTE COMPRESSION Haoran Hong Eduardo Pavez Antonio Ortega Ryosuke Watanabey Keisuke Nonakay.pdf

共5页,预览1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！

MOTION ESTIMATION AND FILTERED PREDICTION FOR DYNAMIC POINT CLOUD ATTRIBUTE COMPRESSION Haoran Hong Eduardo Pavez Antonio Ortega Ryosuke Watanabey Keisuke Nonakay

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: