text
stringlengths 0
89.3k
|
|---|
2
|
FD
|
Y2ˆSQSE
|
arg min
|
Qiµ3
|
2iX
|
ki1PX
|
j1Qk skjQk1 sk1j2
|
2
|
β
|
2PX
|
j1ˆ sijQi sij2
|
2
|
Yi
|
2jˆ sijQi sij
|
34
|
where Yi
|
2jis the rows 3i2through 3iandjth column
|
ofY2 We denote ˆ sijQi sijasˆ rijand compute the
|
gradient of the Lie algebra ϕicorresponding to Qiby imi
|
tating Eq 16
|
ˆ giµ3
|
2PP
|
j1
|
r1
|
ij
|
2
|
2
|
r0
|
ij
|
2
|
2
|
ϕi
|
PP
|
j1β
|
2ˆ rij2
|
2
|
Yi
|
2jˆ rij
|
ϕi
|
µ3PX
|
j1
|
r1
|
ij
|
ϕiTr1
|
ij r0
|
ij
|
ϕiTr0
|
ij
|
PX
|
j1
|
βˆ rij
|
ϕiTˆ rij ˆ rij
|
ϕiTYi
|
2j
|
PX
|
j1
|
Qi sijTh
|
µ3
|
r1
|
ijr0
|
ij
|
βˆ rijYi
|
2ji
|
35
|
Therefore we can still update Qiusing the TPA optimiza
|
tion algorithm in Sec 61 and only need to adjust the resid
|
ual vector rito satisfy the descent direction
|
After discussing the solution formulas for each variable
|
we give the complete optimization Algorithm 2 Since no
|
closedform solution for updating Qexists another iterative
|
optimization must be embedded in the ADMM algorithm
|
In experiments we found that only 1to10iterations are
|
needed to update QwellSmoking
|
a b c d Eating2 Smoking1 Directions
|
e Waiting2
|
Figure 7 Visualization of the 3D shapes in the H3WB NRSfM dataset The first row shows the pictures in the original H36M dataset and
|
the other rows display the GT 3D shapes in each sequence
|
72 Model with Occlusion
|
In realworld scenes the captured images are often ob
|
scured and it is difficult to observe all the keypoints in each
|
frame Assume oiR1P i 1 Fis the mask vec
|
tors and oij 1 ifjth point in the ith frame is visible
|
otherwise 0 Then to solve the occlusion problem we intro
|
duce the mask matrix OR2FPand correct the data term
|
constraint Fas follows
|
µ1
|
2OWΠS2
|
F 36where O 12o112oFR2FP Since the
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.