Datasets:
tensorkelechi
/
arxiv-cstext

Dataset card Data Studio Files Community
1
text
stringlengths
0
89.3k
βM2
iΠT
iΠi T2I3
vecSi
vecµ1
βΠT
iWiM2
iRT
piSiT1
βRT
piYi
3T
43
The update formulas for the other optimization variables in
Model 38 remain unchanged
Orig Sequence
ˆSΛ SProxy ShapesStretch in Mocap
DegenerateSmoothingLow Rank
DoF
Deformation
lowrank friendly Update Nearly Rigid Part
Update
NonRigid
PartFigure 8 Illustration of the proxy shape construction and shape
update principle
8 Supplement to Spatiallyvariant Modeling
In this section We provide an additional explanation of the
construction of the proxy shape and how it plays a role in
optimization refer to Fig 8
1 Definition of proxy shape We divide the nonrigid
object into two regions with different deformation degrees
through frequency domain analysis of 3D trajectories Then
by spatial weighting we merge the single super point
degenerated from NonRigid part with Nearly Rigid
points to form the proxy shapes ieSˆSΛ The weight
matrix Λdefined by the feature mapping ϕEq 12 in
them main text is rankdeficient when αr1 so the re
sultant proxy sequence Shas fewer degrees of freedom than
ˆSand more satisfies the lowrank constraint
2 How SWNN works We enforce the lowrank constraint
on proxy shapes and update new shapes by Eq 27 The
newˆSis mainly composed of Safter lowrank regulariza
tion and Safter smoothing regularization Since Λis rank
deficient the lowrank constrained Safter inverse transfor
mation only retains the structural information of nearly rigid
part Therefore ˆSupdates nearly rigid part mainly through
lowrank and nonrigid part is finetuned by S
Our method combines lowrank and smoothing con
straints through spatial weighting which improves the
accuracy of the reconstruction by avoiding the over
penalization of localized drastic deformations with low
rank constraint To verify the validity of the combination
we designed the comparison experiment presented in Fig 9
We compare the reconstruction error of Eq 20 with meth
ods using only lowrank or smoothing constraints on the
NRSfM Challenge dataset For the lowrankonly method
we compared with RBMM 26 which is an improvement
on the priorfree classical method BMM 11 For the
approach using only the smooth prior we removed the low
rank constraint term from Eq 20 and compared with itFigure 9 Comparison of the reconstruction errors of our model
with methods using only lowrank or smoothing constraints on the
NRSfM Challenge dataset
ie
min
SSQLµ1
2WΠS2
Fµ2
2F1X
i1
QiSiQi1Si1
2
F
β
2
SRpS
2
FD
Y1SRpSE
44
As shown in Fig 9 the reconstruction error of our method
is lower than that of the lowrankonlysmoothonly method
on all types of deformation and there is a significant im
provement on Articul Balloon and Tearing
Our approach can effectively couple lowrank and
smoothing constraints to improve the stability of the algo
rithm over different types of deformations Moreover sta
tistical prior methods tend to be sensitive to the statistical
properties of the data which are also related to the com
plexity of the corresponding object deformation In other
words it is not enough to only mine the smooth prior to
complement the lowrank constraints so finding constraints