tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
Explicit deforming 3D sequence modeling is another dif
ficulty in NRSfM The seminal work by Dai et al 11 re
covers the 3D shape sequence using only the lowrank con
straint Although the lowrank constraint is valid in most
cases it is insufficient to model the objects with severe
spatiallyvariant deformation 27 Kumar et al 26 27
improve the estimation of lowrank structures but still use
global lowrank modeling Some methods 19 48 tried to
introduce additional combinations of linear basis to model
drastic deformation but the lack of effective constraints
makes the recovery of deformation basis difficult
In this paper we propose to resolve the above issues
in camera motion and deforming shape recovery from a
spatialtemporal modeling perspective The object of our
study is the 2D observation sequences and this is a com
mon setup for traditional NRSfM 2 15 17 31 41 Ex
isting methods do not utilize sequence information well
which can help us deal with the above problems effectively
First by digging into the temporal smoothing property
of nonrigid sequences we propose a novel Temporally
smooth Procrustean Alignment TPA module that corrects
camera motion through 3D shape alignment The TPA
module can reduce the effects of rotation ambiguity as
GPA by exploiting the temporal similarity between consec
utive 3D shapes and it also eliminates the need for ad
ditional modeling of reference shapes as the GPAbased
methods 30 41 cf Fig 1 Second to tackle the drastic
spatiallyvariant deformations in realworld 3D shape se
quences we propose to use the spatial information obtained
by analyzing the trajectory space to segment the regions
with different levels of nonrigid deformation Afterward
we propose a SpatialWeighted Nuclear Norm SWNN op
timization to improve the adaptation performance of low
rank constraint for severe spatiallyvariant deformationOur main contributions are summarized as
We propose a Temporallysmooth Procrustean Alignment
TPA module to estimate 3D deforming shapes and ad
just camera motion by aligning the 3D shape sequence
consecutively The TPA module is more conductive to
nonisotropic deformation modeling by remedying the re
quirement of mean shape as alignment reference
We contribute a spatialweighted approach to enforce
the lowrank constraint adaptively at different locations
which can better accommodate severe spatiallyvariant
deformation than global lowrank modeling
We develop a unified spatialtemporal modeling frame
work for NRSfM Extensive experiments across datasets
demonstrate the superiority of our approach Ablation
studies show the effectiveness of each algorithm module
2 Related Work
21 Matrix Factorization based NRSfM
Since Bregler et al 8 first applied the factorization frame
work in nonrigid reconstruction researchers begun to shift
their focus from prelearned modeling methods 4 5 to
the optimization methods such as using Metric Projec
tion 36 37 deformable surface modeling 43 46 and
some priors eg motion states 20 shape priors 12 13
DCT trajectory basis 2 17 etc However Xiao et al 47
pointed out that unlike the rigid version of factorization
framework 44 there is an inherent ambiguity in the solu
tion process using only the orthographic constraint which
would lead to nonunique shape basis and corresponding
coefficients But Akhter et al 3 later proved that inherent
ambiguity does not affect obtaining a unique result Based
on this Dai et al 11 proposed a new pipeline that em
ploys lowrank constraint on camera motion and shape esti
mation combined with the solution technique of Brand 7
this method successfully achieves an unambiguous shape
estimation using the factorization framework without intro
ducing an overly strong prior
Subsequent researchers have derived a number of dif
ferent solution methods based on the application of low
rank constraints in the factorization framework Kumar et
al 26 27 noted that the reconstruction accuracy of 3D
shapes can be improved by preserving the main algebraic
features of the matrix in the nuclear norm optimization
process In addition 1 28 49 model 3D shapes using
union of subspace constraint 10 32 model spatial struc
tures using consensus assumptions and some work repre
sent nonrigid 3D surfaces using manifolds29 39 40 The
factorization framework has been shown to be effective in
NRSfM but a number of issues remain The accuracy of the
unique shape obtained under the existing pipeline is com
promised by the inherent ambiguity and global lowrank
modeling does not well tackle the reconstruction of severely
deformed objects And this paper intends to fill those gaps22 Procrustean Alignment for NRSfM
Parallel to the development of the factorization framework
the alignment framework chose to reduce the impact of the
inherent ambiguity on reconstruction by avoiding estimat
ing motion Torresani et al 45 used Gaussian distribution
as a prior to represent nonrigid shapes and proposed an Ex
pectation Maximization EM solving framework To sepa
rate rigid motion and nonrigid deformation Lee et al 30
introduced the General Procrustean Analysis GPA algo
rithm to construct a special Procrustean Normalized Dis
tribution PND that effectively represents the data proper
ties of nonrigid shapes Lee et al 31 added the hidden
Markov process to PND strengthening the temporal depen
dence in the shape sequence Park et al 41 designed a
more extensible regression framework based on PND and
applied it to deep learning method 42 The separation op

Explicit deforming 3D sequence modeling is another dif

ficulty in NRSfM The seminal work by Dai et al 11 re

covers the 3D shape sequence using only the lowrank con

straint Although the lowrank constraint is valid in most

cases it is insufficient to model the objects with severe

spatiallyvariant deformation 27 Kumar et al 26 27

improve the estimation of lowrank structures but still use

global lowrank modeling Some methods 19 48 tried to

introduce additional combinations of linear basis to model

drastic deformation but the lack of effective constraints

makes the recovery of deformation basis difficult

In this paper we propose to resolve the above issues

in camera motion and deforming shape recovery from a

spatialtemporal modeling perspective The object of our

study is the 2D observation sequences and this is a com

mon setup for traditional NRSfM 2 15 17 31 41 Ex

isting methods do not utilize sequence information well

which can help us deal with the above problems effectively

First by digging into the temporal smoothing property

of nonrigid sequences we propose a novel Temporally

smooth Procrustean Alignment TPA module that corrects

camera motion through 3D shape alignment The TPA

module can reduce the effects of rotation ambiguity as

GPA by exploiting the temporal similarity between consec

utive 3D shapes and it also eliminates the need for ad

ditional modeling of reference shapes as the GPAbased

methods 30 41 cf Fig 1 Second to tackle the drastic

spatiallyvariant deformations in realworld 3D shape se

quences we propose to use the spatial information obtained

by analyzing the trajectory space to segment the regions

with different levels of nonrigid deformation Afterward

we propose a SpatialWeighted Nuclear Norm SWNN op

timization to improve the adaptation performance of low

rank constraint for severe spatiallyvariant deformationOur main contributions are summarized as

We propose a Temporallysmooth Procrustean Alignment

TPA module to estimate 3D deforming shapes and ad

just camera motion by aligning the 3D shape sequence

consecutively The TPA module is more conductive to

nonisotropic deformation modeling by remedying the re

quirement of mean shape as alignment reference

We contribute a spatialweighted approach to enforce

the lowrank constraint adaptively at different locations

which can better accommodate severe spatiallyvariant

deformation than global lowrank modeling

We develop a unified spatialtemporal modeling frame

work for NRSfM Extensive experiments across datasets

demonstrate the superiority of our approach Ablation

studies show the effectiveness of each algorithm module

2 Related Work

21 Matrix Factorization based NRSfM

Since Bregler et al 8 first applied the factorization frame

work in nonrigid reconstruction researchers begun to shift

their focus from prelearned modeling methods 4 5 to

the optimization methods such as using Metric Projec

tion 36 37 deformable surface modeling 43 46 and

some priors eg motion states 20 shape priors 12 13

DCT trajectory basis 2 17 etc However Xiao et al 47

pointed out that unlike the rigid version of factorization

framework 44 there is an inherent ambiguity in the solu

tion process using only the orthographic constraint which

would lead to nonunique shape basis and corresponding

coefficients But Akhter et al 3 later proved that inherent

ambiguity does not affect obtaining a unique result Based

on this Dai et al 11 proposed a new pipeline that em

ploys lowrank constraint on camera motion and shape esti

mation combined with the solution technique of Brand 7

this method successfully achieves an unambiguous shape

estimation using the factorization framework without intro

ducing an overly strong prior

Subsequent researchers have derived a number of dif

ferent solution methods based on the application of low

rank constraints in the factorization framework Kumar et

al 26 27 noted that the reconstruction accuracy of 3D

shapes can be improved by preserving the main algebraic

features of the matrix in the nuclear norm optimization

process In addition 1 28 49 model 3D shapes using

union of subspace constraint 10 32 model spatial struc

tures using consensus assumptions and some work repre

sent nonrigid 3D surfaces using manifolds29 39 40 The

factorization framework has been shown to be effective in

NRSfM but a number of issues remain The accuracy of the

unique shape obtained under the existing pipeline is com

promised by the inherent ambiguity and global lowrank

modeling does not well tackle the reconstruction of severely

deformed objects And this paper intends to fill those gaps22 Procrustean Alignment for NRSfM

Parallel to the development of the factorization framework

the alignment framework chose to reduce the impact of the

inherent ambiguity on reconstruction by avoiding estimat

ing motion Torresani et al 45 used Gaussian distribution

as a prior to represent nonrigid shapes and proposed an Ex

pectation Maximization EM solving framework To sepa

rate rigid motion and nonrigid deformation Lee et al 30

introduced the General Procrustean Analysis GPA algo

rithm to construct a special Procrustean Normalized Dis

tribution PND that effectively represents the data proper

ties of nonrigid shapes Lee et al 31 added the hidden

Markov process to PND strengthening the temporal depen

dence in the shape sequence Park et al 41 designed a

more extensible regression framework based on PND and

applied it to deep learning method 42 The separation op