Datasets:
tensorkelechi
/
arxiv-cstext

Dataset card Data Studio Files Community
1
text
stringlengths
0
89.3k
a2
a3
0a3a2
a3 0a1
a2a1 0
17
Similarly when i 1or F the gradient gican be calcu
lated as
gi
PP
j1
QisijTr0
ij i 1
PP
j1
QisijTr1
ij i F18
The gradient gican be expressed as the product of the Jaco
bian matrix and the residual iegiJT
iriJiR3P3
is the Jacobian matrix as follows
Ji
Qisi1QisiPT 19
where riis the column vector stacked from residuals r1
ij
r0
ijP
j1orrm
ijP
j1 After obtaining the Jacobian ma
trix we can compute the approximation of the secondorder
Hessian matrix as HiJT
iJiand use numerical optimiza
tion methods such as the LevenbergMarquardt LM algo
rithm 33 to solve for the rotation matrix QdiagQi
62 Experiments for TPA
In this subsection we verify the effectiveness of the TPA
module Lets revisit the setting of the experiments We
randomly sample Lie algebra pnR3Ffrom Gaussian
distribution N001 and then map it to a block diagonal
matrix RdR3F3Fconsisting of 3D rotations using the
Rodrigues rotation formula We define the shape sequence
disrupted by random rotations as Snoise RdSgt where
SgtR3FPis the GT sequence in the world coordinate
We realign the sequence using the TPA and GPA 30 mod
ules and compare their results We performed these experi
ments on the Pickup Yoga Stretch and Drink sequences in
MoCap benchmark and the results are shown in Fig 6Yoga
a b c d
Pickup
Stretch
DrinkFigure 6 Experiments for testing the alignment capability of TPA module a Visualization of the shape sequences Snoise b Visu
alization of the TPAaligned shape sequences c Visualization of the GPAaligned shape sequences d The lowrank and smoothing
properties of GT nonrigid sequences Sgt randomly rotated disrupted sequences Snoise and sequences aligned by TPA or GPAWe centralized all the 3D shapes in the sequence and
plotted them in the same coordinate Column a in Fig 6 is
the disrupted sequence Snoise Columns b and c are se
quences after TPA or GPA alignment respectively Column
d shows the lowrank and smoothing properties of real
sequences Sgt noisy sequences Snoise TPAaligned se
quences and GPAaligned sequences Fig 6 demonstrates
that the TPA module can efficiently align the 3D shape se
quences Compared with GPA the TPAaligned sequences
are more similar to the real 3D shape sequences in terms of
their lowrank and smoothing properties
Nonrigid deformation mixed with rigid motion cannot
be accurately estimated using the lowrank constraint As
manifested in Fig 6 d shape sequences after alignment
by the TPA module have lower nuclear norm and smooth
ness From one hand this demonstrates that the TPA mod
ule can effectively separate rigid motion from nonrigid de
formation From the other hand the properties of the TPA
aligned sequence make it can be better recovered by the
lowrank and smoothing constraints In summary we utilize
the lowrank constraint to recover the TPAaligned shape
sequence rather than directly applying the lowrank con
straint under the coordinate defined by the estimated camera
motion 11 26 27
7 ADMM Model Solution
71 Model without Occlusion
In this section we provide the completed solution of our
proposed method Lets first recall the Lagrangian model to
be solved
min
ΩLµ1
2WΠS2
Fµ2
S
µ3
2F1X
i1
QiSiQi1Si1