Taking a Deeper Look at the Inverse Compositional Algorithm
1. Taking a Deeper Look at the Inverse
Compositional Algorithm
Zhaoyang Lv, Frank Dellaert, James M. Rehg, Andreas Geiger
@denkiwakame
2019/02/23 3D勉強会@関
東
1
2. 発表者について
@denkiwakame
● ~’15 Web application Engineer (part-time)
● ?’15 京都大学 松山研究室(B4?M2)
● ?’17 企業研究所
● ‘17? Software Engineer (CV/GPU)
Interests
● 3D Computer Vision
○ Generalized Camera Calibration [M.Nishimura+,ICCV15] ← 修論
● Machine Learning
○ Graphical Models, MRF optimization
● Engineering
○ GPGPU (CUDA) / SIMD / Distributed Computing
2[M.Nishimura+, ICCV15] A Linear Generalized Camera Calibration from Three Intersecting Reference Planes
8. ICLR 2019 Statistics
● An early overview of ICLR2019 (※ 今年5月 開催予定)
○ https://prlz77.github.io/iclr2019-stats/
● Review Ranking (BA-Net Review Score: 8,7,9)
○ https://chillee.github.io/OpenReviewExplorer/
8
476 posters
1418 submissions
22orals
9. 3D幾何最適化 + Deep が注目されている
● 3D勉強会@関東#1
○ CodeSLAM: Learning a Compact, Optimisable Representation for
Dense Visual SLAM
● 3D勉強会@関東#3
○ BA-Net: Dense Bundle Adjustment Network
9
CVPR’18 Best Paper Honorable Mention
ICLR’19 Accept (Oral = 22/1418)
関連する論文を読みます(また)
10. Taking a Deeper Look at the Inverse
Compositional Algorithm
Zhaoyang Lv, Frank Dellaert, James M. Rehg, Andreas Geiger
@denkiwakame
2019/02/23 3D勉強会@関
東
10
11. 著者情報
● Zhaoyang Lv https://www.cc.gatech.edu/~zlv30/
○ 3D Scene Flow, Optical Flow, Motion Tracking
○ SfM, SLAM
● MPI Autonomous Vision Group Intern の成果(?)
11
12. Dense Image Alignment problem
● 2枚の画像間を対応付けるタスク
12[J.Engel+,ECCV14] LSD-SLAM: Large-Scale Direct Monocular SLAM / MPI Sintel Dataset http://sintel.is.tue.mpg.de/
Optical Flow LSD-SLAM
2D motion や 3D のカメラ運動を推定
warp
13. Lucas-Kanade Algorithm Revisit
● 移動先の画素値が等しいと仮定
13[BD Lucas+,IJCAI81] An iterative image registration technique with an application to stereo vision
first-order Taylor expansion
誤差を1次近似
image I transformed
by warping parameter ξ
original
template
Gauss-Newton 法
Δξを求める
近似ヘッセ行列
Related Work
14. Lucas-Kanade Algorithm Revisit
● 移動先の画素値が等しいと仮定
14[BD Lucas+,IJCAI81] An iterative image registration technique with an application to stereo vision
first-order Taylor expansion
誤差を1次近似
image I transformed
by warping parameter ξ
original
template
Gauss-Newton 法
Δξを求める 微分画像(ξk
で変形)
近似ヘッセ行列 ξk
に依存
毎 iteration で計算
ξで微分
Related Work
15. IC (Inverse Compositional) Algorithm
● Additive vs Compositional
● Forward vs Inverse
15
[BD Lucas+,IJCAI81] An iterative image registration technique with an application to stereo vision
[S. Baker+, IJCV04] Lucas-kanade 20 years on: A unifying framework
additive compositional
Forward Inverse
T
I
Related Work
16. ● Lucas-Kanade [BD Lucas+,IJCAI81] ● IC Algorithm [S.Baker+,IJCV04]
16
[BD Lucas+,IJCAI81] An iterative image registration technique with an application to stereo vision
[S. Baker+, IJCV04] Lucas-kanade 20 years on: A unifying framework
IC (Inverse Compositional) Algorithm
parameter update
objective function
first-order Taylor Expansion
ξk
に非依存
pre-compute できる!
Related Work
17. 外れ値がある場合 - ロバスト推定手法
● Robust M-Estimatior
● Robust Version of IC-Algorithm
17
最小二乗法
誤差が正規分布に従うと仮定
Robust M-Estimator
誤差関数ρの選び方が重要
外れ値に小さな重みを与えるような偶関数
誤差関数ρ によって決まる対角行列
Related Work