Datasets:

tensorkelechi

/

arxiv-cstext

like 0

Nonrigid StructurefromMotion Temporallysmooth Procrustean Alignment

and Spatiallyvariant Deformation Modeling

Jiawei Shi Hui Deng Yuchao Dai

School of Electronics and Information Northwestern Polytechnical University

Shaanxi Key Laboratory of Information Acquisition and Processing

sjw2018 denghui986 mailnwpueducn daiyuchaonwpueducn

Abstract

Even though Nonrigid StructurefromMotion NRSfM

has been extensively studied and great progress has been

made there are still key challenges that hinder their broad

realworld applications 1 the inherent motionrotation

ambiguity requires either explicit camera motion recovery

with extra constraint or complex Procrustean Alignment 2

existing lowrank modeling of the global shape can over

penalize drastic deformations in the 3D shape sequence

This paper proposes to resolve the above issues from a

spatialtemporal modeling perspective First we propose

a novel Temporallysmooth Procrustean Alignment module

that estimates 3D deforming shapes and adjusts the cam

era motion by aligning the 3D shape sequence consecu

tively Our new alignment module remedies the requirement

of complex reference 3D shape during alignment which

is more conductive to nonisotropic deformation modeling

Second we propose a spatialweighted approach to enforce

thelowrank constraint adaptively at different locations to

accommodate drastic spatiallyvariant deformation recon

struction better Our modeling outperform existing low

rank based methods and extensive experiments across dif

ferent datasets validate the effectiveness of our method1

1 Introduction

Nonrigid structurefrommotion NRSfM targets at jointly

estimating the deforming 3D shapes and the camera mo

tion from the 2D observation sequence which is a classi

cal and longlasting geometric vision problem 8 In con

trast to the recently booming deep learning based solutions

14 24 25 34 42 48 the optimization based solutions do

not rely on training data and can be directly applied to arbi

trary classes such as human animal and deforming surface

Even though considerable progress has been made 11 and

Yuchao Dai is the corresponding author

1Project page httpsnpucvrgithubioTSMNRSfM

212 1 ˆ 1 iik iki Fk K ci G R 1 3 T

kk rank k K GG Camera Pose World

CoordinateiSa

iRCamera Motion

b

c

World

Coordinate

1iRiR

Temporallysmooth

Alignment iQ2

1 1 min

iii iiFQQS QSTPA

Camera

CoordinateiX

Mean Shape

X

2min

iiiT

iiiiss RtXRX t1 XiRProcrustean

Alignment

GPAProcrustean AlignmentCamera MotionFigure 1 Overview of our proposed TPA Module a Cam

era motion can be recovered by orthographic and rank3 con

straints within the matrix factorization framework11 c The

Procrustean alignment framework uses GPA to resolve the rota

tion ambiguity b We propose the TPA module which aligns the

3D shapes of consecutive frames and corrects the camera motion

by the temporal smoothing property

a series of milestones have been achieved 1 27 30 49 the

performance of NRSfM is still far from satisfactory The

challenges mainly lie in 1 the inherent motionrotation am

biguity requires either explicit camera motion recovery with

extra constraint or complex Procrustean Alignment 2 ex

isting global modeling ie lowrank 11 26 27 unionof

subspace 28 49 of the deforming 3D shape cannot handle

drastic spatiallyvariant 3D deformations

Existing NRSfM statistical prior frameworks can be

roughly classified into two categories 1 explicit motion

estimationbased approaches 2 8 11 15 26 27 and 2

motion estimationfree approaches 30 31 41 The for

mer usually employs the matrix factorization frameworkarXiv240504309v1 csCV 7 May 2024and orthographic constraint to solve for camera motion and

3D shapes separately Dai et al 11 introduced the low

rank constraint into the recovery of camera motion Ku

mar et al dig into the correction matrix to improve the

motion estimation results by using the smoothing prior of

camera motions 26 or the fusion of different camera mo

tion sequences 27 However these camera motion estima

tion methods have difficulties in tackling the inherent rota

tionmotion ambiguity problem in NRSfM 47

The latter solves the shapes directly and remedies the

requirement of motion estimation Lee et al 30 31 41

argued that nonrigid shapes can be recovered unambigu

ously under the coordinate with almost no rigid relative

motion This type of framework resolves the rotation am

biguity through General Procrustean Analysis GPA 18

to achieve alignment of the 3D shapes However the

high dependence on the alignment reference shape makes

the model extremely complex 30 41 and tends to over

penalize nonisotropic deformations