Toru Tamaki, Bisser Raytchev, Kazufumi Kaneda, Toshiyuki Amano : "Wstimating a Rotation Matrix R by using higher-order Matrices R^n with Application to Supervised Pose Estimation," Proc. of SPPRA 2010: The Seventh IASTED International Conference on Signal Processing, Pattern Recognition and Applications, pp. 58-64 (2010 02). Innsbruck, Austria, 2010/February/17-19.
SPPRA2010 Estimating a Rotation Matrix R by using higher-order Matrices R^n with Application to Supervised Pose Estimation
1. Estimating a rotation matrix R
by using higher-order matrices Rn
with application to
supervised pose estimation
Toru Tamaki
Bisser Raytchev
Kazufumi Kaneda
Toshiyuki Amano
2. Estimating a rotation matrix R
by using higher-order matrices Rn
with application to
supervised pose estimation
Can Rnestimate R
Toru Tamaki
more accurately Bisser Raytchev
than R ? Kazufumi Kaneda
Toshiyuki Amano
3. To improve estimates… Average!
t1 When measurement is only once…
t2
t
…
…
Average
tn Lease-Squares
How improve?
Measure many times
4. To improve estimates… Average!
t1 When measurement is only once…
t2
t
…
…
Average
tn Lease-Squares
Improve!
Measure with many timers
5. To improve estimates… Average?
t When measurement is only once…
t2
t
…
…
Average
tn Lease-Squares
How improve?
Measure with many timers
6. To improve estimates… Average.
t When measurement is only once…
t2
t
…
…
tn
Improve!
Measure with many timers
But, What is it?
7. Our problem: Pose estimation
Pose parameters
R 3x3 rotation matrix
t 3D translation
Regression:
image Appearance-based /
Pose
parameters View-based pose estimation
R Parametric Eigenspace (Murase et al., 1995)
linear regression (Okatani et al., 2000)
kernel CCA (Melzer et al., 2003)
SV regression (Ando et al., 2005)
Manifold learning, and others
R (Rothganger et al., 2006) (Lowe, 2004) (Ferrari et
al., 2006) (Kushal et al., 2006) (Viksten, 2009)
12. Principle of EDM
Transmitter Receiver
Dist
¸1
¸2
µ2
µ1
device target
Use
•Longer wavelength ¸1 first, for a rough phase estimate µ1
•Shorter wavelength ¸2 next, for a fine phase estimate µ2
13. Our concept
New
Pose Rotation
image Examples
vector matrix axis angle
1
p1 R ! µ 29 [deg] 210 [deg]
! Polar Eigen
Decomp. Decomp.
µ
62 [deg] 420 [deg]
2 p2 R2 ! 2µ =60 [deg]
! Div by 2 Div by 2
31 [deg] 210 [deg]
2µ ! µ
30 [deg] 210 [deg]
14. Measurements Simulation 1
Pose Rotation
vector matrix axis angle
R R p1 +noise R !1 µ1
… Polar Eigen
…
…
…
…
Decomp. Decomp.
µ
R8 p8 +noise R8 !8 µ8
! R
15. Measurements
Simulation 2
Pose Rotation
vector matrix axis angle
R +noise R p1 R !1 µ1
Polar Eigen
…
…
…
…
…
Decomp. Decomp.
µ
R8 p8 R8 !8 µ8
! R
Error doesn’t change…
No free lunch!
16. •Linear regression
Experimental results
•Training with
Pose Rotation
images and poses
vector matrix axis angle
1
p1 R !1 µ1
Polar Eigen
…
…
…
…
Decomp. Decomp.
µ
8
p8 R8 !8 µ8
images
! R
17. Summary
• Improve estimates of a pose R with
many measurements R1, R2, …, R8
• Simulations and experimental results
shows that the concept works!