tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
eration of the alignment framework has a significant effect
on mitigating the rotation ambiguity than most of factor
ization methods but on the other hand the pure alignment
framework fails to capitalize on the advantages of the fac
torization framework as well Therefore this paper aims to
combine the advantages of both to obtain better results
3 Method
In this section we introduce our spatialtemporal modeling
solution for NRSfM We first recap the definition of two
classical models in NRSfM Then we improve the model
ing of nonrigid deformations from both temporal and spa
tial perspectives ie temporallysmooth Procrustean align
ment and spatialweighted nonrigid deformation modeling
31 Problem Formulation and Optimization
NRSfM aims to recover 3D deforming shape sequence S
S1SFR3FPin the world coordinate and camera
motion RdiagRiR2F3Ffrom 2D measurements
W W1WFR2FPieWRS 8F
andPdenote the number of frames and points The ma
trix factorization model 8 11 44 assumes the nonrigid
3D shapes can be expressed by the linear combination of
Kshape basis BjieSiPK
j1cjBj where cjis the
shape basis coefficients Following this assumption Rand
Scan be estimated by the Singular Value Decomposition
SVD of Wcoupled with orthographic constraint Since
QSO3WiRiQTQSialways holds there is
an inherent rotation ambiguity in the absence of additional
constraints which leads to incorrect estimation of camera
motion and 3D shapes 11 26 27 have proposed a num
ber of effective priorfree methods to recover camera mo
tion and shape sequence but these methods can not resolve
the rotation ambiguity problem theoretically
Different from the above methods another class of
methods30 41 estimates alignment rotations by Procrustean Alignment to substitute the recovery of camera
motion The General Procrustean Alignment GPA model
is defined as
Ri arg min
Ri

RiSiTS

RiSO3 1
where TI1
P11Tis the translation elimination matrix S
is the mean shape used as the reference shape during align
ment and Siis the 3D shape in camera coordinate Such
methods effectively resolve the rotation ambiguity by sep
arating rigid motion from nonrigid deformation But for
nonisotropic deformations variations along certain direc
tions may markedly influence the mean shape leading to an
exaggerated penalization of shapes during optimization
In addition to the reprojection information Win cam
era coordinate both the factorization framework and Pro
crustean alignment framework are looking for a canonical
coordinate ie world or alignment coordinate to enforce
extra regularization such as lowrank We use SR3FP
to denote the sequence of 3D shapes in the camera coordi
nate and ˆSR3FPis the shape sequence in the canonical
coordinate We rethink these two types of frameworks and
provide a new unified framework
min
SˆSRFS GˆS PRΦ
stˆSRS2
where Fdenotes the data term such as reprojection
error and Gdenotes the regularization term such as
lowrank or smoothing Prepresents the transforma
tion refinement module Ris the transformation between
the camera and canonical coordinates Φis the optimization
parameters Rdenotes pure 3D rotation transformation
while we have removed scale and translation from measure
ment matrix Wby regularization and centralization as 11
When Ris computed in advance and fixed during optimiza
tion the model 2 degenerates to priorfree methods such
as 11 26 27 If the optimization goal Pis set to GPA
1 and Φis defined as SS model 2 degenerates into
the framework in 41
32 Temporallysmooth Procrustean Alignment
Under the matrix factorization framework the 3D recon
struction results will be affected by camera motion esti
mation as incorrect camera motion estimation may invali
date the underlying lowrank assumption The Procrustean
alignment methods make the aligned sequence satisfy the
lowrank constraint by minimizing the rigid motion be
tween 3D shapes Therefore such methods can obtain im
proved results when they begin with good initialization
In the real world the nonrigid objects deform contin
uously and smoothly Therefore the low rank as well asFrequency3D Trajectories
Scaled Periodogram
jth Joint1ˆF
tjis
DFT
12Deformation
Frequency
f
j k f d m
Segmentation ResultSortNearlyRigid
vs
NonRigid
ra Process for nearly rigid region segmentation

eration of the alignment framework has a significant effect

on mitigating the rotation ambiguity than most of factor

ization methods but on the other hand the pure alignment

framework fails to capitalize on the advantages of the fac

torization framework as well Therefore this paper aims to

combine the advantages of both to obtain better results

3 Method

In this section we introduce our spatialtemporal modeling

solution for NRSfM We first recap the definition of two

classical models in NRSfM Then we improve the model

ing of nonrigid deformations from both temporal and spa

tial perspectives ie temporallysmooth Procrustean align

ment and spatialweighted nonrigid deformation modeling

31 Problem Formulation and Optimization

NRSfM aims to recover 3D deforming shape sequence S

S1SFR3FPin the world coordinate and camera

motion RdiagRiR2F3Ffrom 2D measurements

W W1WFR2FPieWRS 8F

andPdenote the number of frames and points The ma

trix factorization model 8 11 44 assumes the nonrigid

3D shapes can be expressed by the linear combination of

Kshape basis BjieSiPK

j1cjBj where cjis the

shape basis coefficients Following this assumption Rand

Scan be estimated by the Singular Value Decomposition

SVD of Wcoupled with orthographic constraint Since

QSO3WiRiQTQSialways holds there is

an inherent rotation ambiguity in the absence of additional

constraints which leads to incorrect estimation of camera

motion and 3D shapes 11 26 27 have proposed a num

ber of effective priorfree methods to recover camera mo

tion and shape sequence but these methods can not resolve

the rotation ambiguity problem theoretically

Different from the above methods another class of

methods30 41 estimates alignment rotations by Procrustean Alignment to substitute the recovery of camera

motion The General Procrustean Alignment GPA model

is defined as

Ri arg min

Ri

RiSiTS

RiSO3 1

where TI1

P11Tis the translation elimination matrix S

is the mean shape used as the reference shape during align

ment and Siis the 3D shape in camera coordinate Such

methods effectively resolve the rotation ambiguity by sep

arating rigid motion from nonrigid deformation But for

nonisotropic deformations variations along certain direc

tions may markedly influence the mean shape leading to an

exaggerated penalization of shapes during optimization

In addition to the reprojection information Win cam

era coordinate both the factorization framework and Pro

crustean alignment framework are looking for a canonical

coordinate ie world or alignment coordinate to enforce

extra regularization such as lowrank We use SR3FP

to denote the sequence of 3D shapes in the camera coordi

nate and ˆSR3FPis the shape sequence in the canonical

coordinate We rethink these two types of frameworks and

provide a new unified framework

min

SˆSRFS GˆS PRΦ

stˆSRS2

where Fdenotes the data term such as reprojection

error and Gdenotes the regularization term such as

lowrank or smoothing Prepresents the transforma

tion refinement module Ris the transformation between

the camera and canonical coordinates Φis the optimization

parameters Rdenotes pure 3D rotation transformation

while we have removed scale and translation from measure

ment matrix Wby regularization and centralization as 11

When Ris computed in advance and fixed during optimiza

tion the model 2 degenerates to priorfree methods such

as 11 26 27 If the optimization goal Pis set to GPA

1 and Φis defined as SS model 2 degenerates into

the framework in 41

32 Temporallysmooth Procrustean Alignment

Under the matrix factorization framework the 3D recon

struction results will be affected by camera motion esti

mation as incorrect camera motion estimation may invali

date the underlying lowrank assumption The Procrustean

alignment methods make the aligned sequence satisfy the

lowrank constraint by minimizing the rigid motion be

tween 3D shapes Therefore such methods can obtain im

proved results when they begin with good initialization

In the real world the nonrigid objects deform contin

uously and smoothly Therefore the low rank as well asFrequency3D Trajectories

Scaled Periodogram

jth Joint1ˆF

tjis

DFT

12Deformation

Frequency

f

j k f d m

Segmentation ResultSortNearlyRigid

vs

NonRigid

ra Process for nearly rigid region segmentation