Translation Synchronization via Truncated Least Squares Xiangru Huang 1 * Zhenxiao Liang 2 * Chandrajit Bajaj 1 Qixing Huang 1 1 Department of Computer Science University of Texas at Austin 2 Tsinghua University NIPS, 2017 Xiangru Huang *, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 1 / 20
24
Embed
Translation Synchronization via Truncated Least Squaresxrhuang/slides/TranSyncSpotlight_NIPS17.pdf · Translation Synchronization via Truncated Least Squares Xiangru Huang1* Zhenxiao
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Translation Synchronizationvia Truncated Least Squares
1Department of Computer ScienceUniversity of Texas at Austin
2Tsinghua University
NIPS, 2017
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 1 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 2 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
µ
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 3 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
µ
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 4 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
µ
mean
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 5 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 6 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 7 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
mean
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 8 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 9 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
1. Delete Sample tj if |tj −mean|>ǫ
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 10 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
1. Delete Sample tj if |tj −mean|>ǫ
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 11 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
1. Delete Sample tj if |tj −mean|>ǫc1
2. Recompute mean and Shrink threshold
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 12 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
1. Delete Sample tj if |tj −mean|>ǫc2
2. Recompute mean and Shrink threshold
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 13 / 20
From a Simple Example
U[µ− σ, µ+ σ]
σ−σ
Outliers
µ
meanmedian
1. Delete Sample tj if |tj −mean|>ǫc3
2. Recompute mean and Shrink threshold
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 14 / 20
Translation Synchronization
Ground Truth {xi}Relative measurements tij = xi − xj + noise ∀i , j ∈ E
Algorithm: iteratively update x and E .Can be applied to Pairwise Ranking, Joint Alignment of point clouds,and etc.
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 15 / 20
Translation Synchronization
Ground Truth {xi}Relative measurements tij = xi − xj + noise ∀i , j ∈ EAlgorithm: iteratively update x and E .
Can be applied to Pairwise Ranking, Joint Alignment of point clouds,and etc.
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 15 / 20
Translation Synchronization
Ground Truth {xi}Relative measurements tij = xi − xj + noise ∀i , j ∈ EAlgorithm: iteratively update x and E .Can be applied to Pairwise Ranking, Joint Alignment of point clouds,and etc.
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 15 / 20
Exact Recovery
Biased Noise Model (Unbounded Outliers):
tij =
{xgti − xgtj + U[−σ, σ] with probability p
Any real number with probability 1− p(1)
For some constants p, q only depend on graph structure, duringoptimization we have
‖x (k) − xgt‖∞ ≤ qσ + 2pεck−1
and eventually we’ll reach an x̂
‖x̂ − xgt‖∞ ≤2p + cq
c − 4pσ
where the RHS is independent of ε.
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 16 / 20
Exact Recovery
Biased Noise Model (Unbounded Outliers):
tij =
{xgti − xgtj + U[−σ, σ] with probability p
Any real number with probability 1− p(1)
For some constants p, q only depend on graph structure, duringoptimization we have
‖x (k) − xgt‖∞ ≤ qσ + 2pεck−1
and eventually we’ll reach an x̂
‖x̂ − xgt‖∞ ≤2p + cq
c − 4pσ
where the RHS is independent of ε.
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 16 / 20
Randomized Case
Biased Noise Model:
tij =
{xgti − xgtj + U[−σ, σ] with probability p
xgti − xgtj + U[−a, b] with probability 1− p(2)
TheoremThere exists a constant c so that if p > c/
√log(n), then w.h.p,
‖x (k) − xgt‖∞ ≤ (1− p/2)k(b − a),
∀ k = 0, · · · , [− log(b + a
2σ)/log(1− p/2)].
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 17 / 20
Randomized Case
Biased Noise Model:
tij =
{xgti − xgtj + U[−σ, σ] with probability p
xgti − xgtj + U[−a, b] with probability 1− p(2)
TheoremThere exists a constant c so that if p > c/
√log(n), then w.h.p,
‖x (k) − xgt‖∞ ≤ (1− p/2)k(b − a),
∀ k = 0, · · · , [− log(b + a
2σ)/log(1− p/2)].
Xiangru Huang*, Zhenxiao Liang* , Chandrajit Bajaj , Qixing Huang Translation Synchronization via Truncated Least Squares 17 / 20