its applications to rotation synchronization
play

Its Applications to Rotation Synchronization Yunpeng Shi Joint work - PowerPoint PPT Presentation

Message Passing Least Squares Framework and Its Applications to Rotation Synchronization Yunpeng Shi Joint work with Prof. Gilad Lerman School of Mathematics University of Minnesota Rotation Synchronization 3 2 1 5 4 Given a graph


  1. Message Passing Least Squares Framework and Its Applications to Rotation Synchronization Yunpeng Shi Joint work with Prof. Gilad Lerman School of Mathematics University of Minnesota

  2. Rotation Synchronization 3 2 1 5 4 Given a graph ๐ป ๐‘œ , ๐น ๐‘œ : = 1,2,3, โ€ฆ , ๐‘œ , ๐น is the set of edges

  3. Rotation Synchronization โˆ— = ? ๐‘† 3 โˆ— = ? ๐‘† 2 โˆ— = ? ๐‘† 1 โˆ— = ? โˆ— = ? ๐‘† 4 ๐‘† 5 โˆ— โˆˆ ๐‘‡๐‘ƒ(3) Each node ๐‘— โˆˆ [๐‘œ] is assigned an unknown ground truth rotation ๐‘† ๐‘—

  4. Rotation Synchronization ๐‘† 23 โˆ— = ? ๐‘† 12 ๐‘† 3 โˆ— = ? ๐‘† 2 โˆ— = ? ๐‘† 1 ๐‘† 25 ๐‘† 35 ๐‘† 14 โˆ— = ? โˆ— = ? ๐‘† 4 ๐‘† 5 ๐‘† 45 โ€ข Each edge ๐‘—๐‘˜ โˆˆ ๐น is given a possibly noisy and corrupted relative rotation ๐‘† ๐‘—๐‘˜ โˆ— = ๐‘† ๐‘— โˆ— ๐‘† ๐‘˜ โˆ—โˆ’1 โ€ข The uncorrupted relative rotation for ๐‘—๐‘˜ โˆˆ ๐น is ๐‘† ๐‘—๐‘˜ โˆ— โ€ข Rotation Synchronization : Estimate ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] from ๐‘† ๐‘—๐‘˜ ๐‘—๐‘˜โˆˆ๐น โˆ— ๐‘† ๐‘—โˆˆ[๐‘œ] for any rotation ๐‘† is also a solution โ€ข ๐‘† ๐‘—

  5. Applications Camera orientation estimation in 3D reconstruction tasks: Structure from motion (SfM) Simultaneous localization and mapping (SLAM) Demonstration by Carl Olsson Demonstration by Raรบl Mur-Artal

  6. Adversarial Corruption Model โˆ— : = ๐‘† ๐‘— โˆ— ๐‘† ๐‘˜ โˆ—โˆ’1 , ๐‘† ๐‘—๐‘˜ ๐‘—๐‘˜ โˆˆ ๐น ๐‘• (good edges) ๐‘† ๐‘—๐‘˜ = แ‰ เทจ ๐‘† ๐‘—๐‘˜ , ๐‘—๐‘˜ โˆˆ ๐น ๐‘ (bad edges)

  7. Adversarial Corruption Model โˆ— : = ๐‘† ๐‘— โˆ— ๐‘† ๐‘˜ โˆ—โˆ’1 , ๐‘† ๐‘—๐‘˜ ๐‘—๐‘˜ โˆˆ ๐น ๐‘• (good edges) ๐‘† ๐‘—๐‘˜ = แ‰ เทจ ๐‘† ๐‘—๐‘˜ , ๐‘—๐‘˜ โˆˆ ๐น ๐‘ (bad edges) โˆ— โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) Corruption Level ๐‘ก ๐‘—๐‘˜ Commonly, ๐‘’ is the geodesic distance on ๐‘‡๐‘ƒ(3)

  8. Least Squares Solvers โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) เท minimize ๐‘˜ ๐‘—๐‘˜โˆˆ๐น The most common approximate solution is the Lie algebraic averaging

  9. Robust Solvers: ๐‘š ๐‘ž minimization ( 0 < ๐‘ž โ‰ค 1 ) โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘’ ๐‘ž (๐‘† ๐‘— ๐‘† ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) เท minimize ๐‘˜ ๐‘—๐‘˜โˆˆ๐น

  10. โˆ’1 , ๐‘† ๐‘—๐‘˜ ) over ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] in SO(3)? How to minimize ฯƒ ๐‘—๐‘˜โˆˆ๐น ๐‘’ ๐‘ž (๐‘† ๐‘— ๐‘† ๐‘˜

  11. โˆ’1 , ๐‘† ๐‘—๐‘˜ ) over ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] in SO(3)? How to minimize ฯƒ ๐‘—๐‘˜โˆˆ๐น ๐‘’ ๐‘ž (๐‘† ๐‘— ๐‘† ๐‘˜ Iteratively Reweighted Least Squares (IRLS): โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ฯƒ ๐‘—๐‘˜โˆˆ๐น ๐‘ฅ ๐‘—๐‘˜,๐‘ข ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘˜ ๐‘† ๐‘—,๐‘ข ๐‘—โˆˆ[๐‘œ] = argmin ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘  ๐‘—๐‘˜,๐‘ข = ๐‘’(๐‘† ๐‘—,๐‘ข ๐‘† ๐‘˜,๐‘ข ๐‘žโˆ’2 ๐‘ฅ ๐‘—๐‘˜,๐‘ข+1 = ๐‘  ๐‘—๐‘˜,๐‘ข

  12. โˆ’1 , ๐‘† ๐‘—๐‘˜ ) over ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] in SO(3)? How to minimize ฯƒ ๐‘—๐‘˜โˆˆ๐น ๐‘’ ๐‘ž (๐‘† ๐‘— ๐‘† ๐‘˜ Iteratively Reweighted Least Squares (IRLS): โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ฯƒ ๐‘—๐‘˜โˆˆ๐น ๐‘ฅ ๐‘—๐‘˜,๐‘ข ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘˜ ๐‘† ๐‘—,๐‘ข ๐‘—โˆˆ[๐‘œ] = argmin ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘  ๐‘—๐‘˜,๐‘ข = ๐‘’(๐‘† ๐‘—,๐‘ข ๐‘† ๐‘˜,๐‘ข ๐‘žโˆ’2 ๐‘ฅ ๐‘—๐‘˜,๐‘ข+1 = ๐‘  ๐‘—๐‘˜,๐‘ข 2โˆ’๐‘ž โˆ— โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) and ๐‘ฅ ๐‘—๐‘˜,๐‘ข+1 โ‰ˆ 1 Ideally, ๐‘  ๐‘—๐‘˜,๐‘ข โ‰ˆ ๐‘ก ๐‘—๐‘˜ concentrates on ๐น ๐‘• โˆ— ๐‘ก ๐‘—๐‘˜

  13. Issue 1: Over-Aggressive Reweighting โˆ— and thus ๐‘  ๐‘—๐‘˜,๐‘ข โ‰‰ ๐‘ก ๐‘—๐‘˜ โˆ— โ€ข Under severe corruption, ๐‘† ๐‘—,๐‘ข โ‰‰ ๐‘† ๐‘— โ€ข In certain cases, ๐‘  ๐‘—๐‘˜,๐‘ข โ‰ˆ 0 for ๐‘—๐‘˜ โˆˆ ๐น ๐‘ , and thus 2โˆ’๐‘ž 1 ๐‘ฅ ๐‘—๐‘˜,๐‘ข+1 = can be extremely high for ๐‘—๐‘˜ โˆˆ ๐น ๐‘ ๐‘  ๐‘—๐‘˜,๐‘ข

  14. Issue 2: Poor Least Squares Solution โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘ฅ ๐‘—๐‘˜,๐‘ข ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘˜ ๐‘† ๐‘—,๐‘ข ๐‘—โˆˆ[๐‘œ] = argmin เท ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) ๐‘—๐‘˜โˆˆ๐น

  15. Issue 2: Poor Least Squares Solution โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘ฅ ๐‘—๐‘˜,๐‘ข ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘˜ ๐‘† ๐‘—,๐‘ข ๐‘—โˆˆ[๐‘œ] = argmin เท ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) ๐‘—๐‘˜โˆˆ๐น Lie algebraic representation ๐œ• ๐‘— ๐‘† ๐‘— Riemannian manifold of SO(3) ๐‘† ๐‘—,๐‘ขโˆ’1

  16. Issue 2: Poor Least Squares Solution โˆ’1 , ๐‘† ๐‘—๐‘˜ ) ๐‘ฅ ๐‘—๐‘˜,๐‘ข ๐‘’ 2 (๐‘† ๐‘— ๐‘† ๐‘˜ ๐‘† ๐‘—,๐‘ข ๐‘—โˆˆ[๐‘œ] = argmin เท ๐‘† ๐‘— ๐‘—โˆˆ[๐‘œ] โŠ‚๐‘‡๐‘ƒ(3) ๐‘—๐‘˜โˆˆ๐น ๐œ• ๐‘— The Lie Algebraic Averaging uses the approximation โˆ’1 , ๐‘† ๐‘—๐‘˜ โ‰ˆ ิก ๐œ• ๐‘— โˆ’ ๐‘’ ๐‘† ๐‘— ๐‘† ๐œ• ๐‘˜ โˆ’ ๐œ• ๐‘—๐‘˜ เธฎ ๐‘† ๐‘— ๐‘˜ 2 โˆ’1 ๐‘† ๐‘— ) where ๐œ• ๐‘— = log( ๐‘† ๐‘—,๐‘ขโˆ’1 ๐‘† ๐‘—,๐‘ขโˆ’1 โˆ’1 ๐‘† ๐‘—๐‘˜ ๐‘† and ๐œ• ๐‘—๐‘˜ = log( ๐‘† ๐‘—,๐‘ขโˆ’1 ๐‘˜,๐‘ขโˆ’1 ) โˆ— and ๐‘† ๐‘—๐‘˜ โ‰ˆ ๐‘† ๐‘— โˆ— ๐‘† ๐‘˜ โˆ—โˆ’1 The approximation is valid only when ๐‘† ๐‘— โ‰ˆ ๐‘† ๐‘—

  17. โˆ— without knowing ๐‘† ๐‘— โˆ— and ๐‘† ๐‘˜ โˆ— ? How to accurately estimate ๐‘ก ๐‘—๐‘˜

  18. Cycle-Edge Message Passing (CEMP) โ€ข Goal: Estimate corruption level โˆ— โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) ๐‘ก ๐‘—๐‘˜ from 3-cycle inconsistency measure ๐‘’ ๐‘—๐‘˜๐‘™ โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜๐‘™ ๐‘† ๐‘™๐‘— , ๐ฝ)

  19. Cycle-Edge Message Passing (CEMP) โ€ข Goal: Estimate corruption level โˆ— โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) ๐‘ก ๐‘—๐‘˜ from 3-cycle inconsistency measure ๐‘’ ๐‘—๐‘˜๐‘™ โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜๐‘™ ๐‘† ๐‘™๐‘— , ๐ฝ) โ€ข For each ๐‘—๐‘˜ โˆˆ ๐น, sample 50 3-cycles ๐‘—๐‘˜๐‘™ and for each cycle compute ๐‘’ ๐‘—๐‘˜๐‘™

  20. Cycle-Edge Message Passing (CEMP) โ€ข Goal: Estimate corruption level โˆ— โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) ๐‘ก ๐‘—๐‘˜ from 3-cycle inconsistency measure ๐‘’ ๐‘—๐‘˜๐‘™ โ‰” ๐‘’(๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜๐‘™ ๐‘† ๐‘™๐‘— , ๐ฝ) โ€ข For each ๐‘—๐‘˜ โˆˆ ๐น, sample 50 3-cycles ๐‘—๐‘˜๐‘™ and for each cycle compute ๐‘’ ๐‘—๐‘˜๐‘™ โ€ข For fixed ๐‘—๐‘˜ and good 3-cycles w.r.t. ๐‘—๐‘˜ (That is, ๐‘—๐‘™, ๐‘˜๐‘™ โˆˆ ๐น ๐‘• ) โˆ— ๐‘† ๐‘™๐‘— โˆ— , ๐ฝ) = ๐‘’(๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜๐‘™ โˆ— ๐‘† ๐‘™๐‘— โˆ— ๐‘† ๐‘—๐‘˜ โˆ— , ๐‘† ๐‘—๐‘˜ โˆ— ) = ๐‘’(๐‘† ๐‘—๐‘˜ , ๐‘† ๐‘—๐‘˜ โˆ— ) = ๐‘ก ๐‘—๐‘˜ โˆ— ๏ธธ ๐‘’ ๐‘—๐‘˜๐‘™ = ๐‘’(๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜๐‘™ = ๐ฝ by cycle consistency

  21. ๐‘’ ๐‘—๐‘˜๐‘™ 1 ๐‘’ ๐‘—๐‘˜๐‘™ 2 ๐‘’ ๐‘—๐‘˜๐‘™ 49 ๐‘’ ๐‘—๐‘˜๐‘™ 50

  22. ๐‘’ ๐‘—๐‘˜๐‘™ 1 ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 1 โˆ— ๐‘’ ๐‘—๐‘˜๐‘™ 2 = ๐‘ก ๐‘—๐‘˜ ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 2 ๐‘—๐‘˜ ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 49 โˆ— ๐‘’ ๐‘—๐‘˜๐‘™ 49 = ๐‘ก ๐‘—๐‘˜ ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 50 ๐‘’ ๐‘—๐‘˜๐‘™ 50

  23. ๐‘’ ๐‘—๐‘˜๐‘™ 1 ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 1 ๐‘’ ๐‘—๐‘˜๐‘™ 2 โˆ— = ๐‘ก ๐‘—๐‘˜ ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 2 โˆ— โ‰ˆ ๐‘ก ๐‘—๐‘˜,๐‘ข+1 : = ๐‘ข ๐‘’ ๐‘—๐‘˜๐‘™ 1 ๐‘—๐‘˜ ๐‘ข ฯƒ ๐‘™ ๐‘ž ๐‘—๐‘˜๐‘™ ๐‘ก ๐‘—๐‘˜ ๐‘Ž ๐‘—๐‘˜ ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ 49 ๐‘’ ๐‘—๐‘˜๐‘™ 49 ๐‘ข ๐‘’ ๐‘—๐‘˜๐‘™ ๐‘ข = ฯƒ ๐‘™ ๐‘ž ๐‘—๐‘˜๐‘™ ๐‘Ž ๐‘—๐‘˜ โˆ— ๐‘ข = ๐‘ก ๐‘—๐‘˜ ๐‘ž ๐‘—๐‘˜๐‘™ 50 ๐‘’ ๐‘—๐‘˜๐‘™ 50

  24. ๐‘ก ๐‘—๐‘™,๐‘ข ๐‘ก ๐‘˜๐‘™,๐‘ข

  25. Prob. that ๐‘—๐‘™ โˆˆ ๐น ๐‘• given ๐‘ก ๐‘—๐‘™,๐‘ข ๐‘“ โˆ’๐›พ ๐‘ข โˆ™ ๐‘ž ๐‘—๐‘™,๐‘ข ๐‘ก ๐‘—๐‘™,๐‘ข ๐‘“ โˆ’๐›พ ๐‘ข โˆ™ ๐‘ž ๐‘˜๐‘™,๐‘ข ๐‘ก ๐‘˜๐‘™,๐‘ข Prob. that j๐‘™ โˆˆ ๐น ๐‘• given ๐‘ก ๐‘˜๐‘™,๐‘ข

  26. Prob. that ๐‘—๐‘™ โˆˆ ๐น ๐‘• given ๐‘ก ๐‘—๐‘™,๐‘ข ๐‘“ โˆ’๐›พ ๐‘ข โˆ™ ๐‘ž ๐‘—๐‘™,๐‘ข ๐‘ก ๐‘—๐‘™,๐‘ข Prob. that ๐‘—๐‘˜๐‘™ is good given {๐‘ก ๐‘๐‘,๐‘ข : ๐‘๐‘ โˆˆ ๐น} ๐‘—๐‘˜๐‘™ ๐‘ข = ๐‘“ โˆ’๐›พ ๐‘ข (๐‘ก ๐‘—๐‘™,๐‘ข +๐‘ก ๐‘˜๐‘™,๐‘ข ) ๐‘ž ๐‘—๐‘˜๐‘™ ๐‘“ โˆ’๐›พ ๐‘ข โˆ™ ๐‘ž ๐‘˜๐‘™,๐‘ข ๐‘ก ๐‘˜๐‘™,๐‘ข Prob. that j๐‘™ โˆˆ ๐น ๐‘• given ๐‘ก ๐‘˜๐‘™,๐‘ข

  27. Cycle-Edge Message Passing (CEMP) โˆ— ): The conditional probability that ๐‘—๐‘˜๐‘™ is good ( ๐‘’ ๐‘—๐‘˜๐‘™ = ๐‘ก ๐‘—๐‘˜ = ๐‘“ โˆ’๐›พ ๐‘ข (๐‘ก ๐‘—๐‘™,๐‘ข +๐‘ก ๐‘˜๐‘™,๐‘ข ) = ๐‘„๐‘ (๐‘’ ๐‘—๐‘˜๐‘™ = ๐‘ก ๐‘—๐‘˜ โˆ— |{๐‘ก ๐‘๐‘,๐‘ข : ๐‘๐‘ โˆˆ ๐น}) ๐‘ข ๐‘ž ๐‘—๐‘˜๐‘™ The estimate of the corruption level: ๐‘ข ๐‘’ ๐‘—๐‘˜๐‘™ = ๐”ฝ(๐‘ก ๐‘—๐‘˜ โˆ— |{๐‘ก ๐‘๐‘,๐‘ข : ๐‘๐‘ โˆˆ ๐น}) 1 ๐‘ข ฯƒ ๐‘™ ๐‘ž ๐‘—๐‘˜๐‘™ ๐‘ก ๐‘—๐‘˜,๐‘ข+1 : = ๐‘Ž ๐‘—๐‘˜

  28. Theory If f 1 โ€ข the maximal ratio of corrupted cycles per edge < 5 โ€ข ๐›พ ๐‘ข increases exponentially with a sufficiently small rate, the hen โ€ข for all ๐‘—๐‘˜ โˆˆ ๐น, ๐‘ก ๐‘—๐‘˜,๐‘ข computed by CEMP linearly and uniformly โˆ— . converges to ๐‘ก ๐‘—๐‘˜

  29. Message Passing Least Squares (MPLS) Init Initialization Run CEMP for ๐‘ˆ iterations Build a weighted graph with edge weights ๐‘ก ๐‘—๐‘˜,๐‘ˆ ๐‘—๐‘˜โˆˆ๐น Find the minimal spanning tree using Primโ€™s Algorithm ize ๐‘† ๐‘—,0 by fixing ๐‘† 1,0 = ๐ฝ and ๐‘† ๐‘—,0 = ๐‘† ๐‘—๐‘˜ ๐‘† ๐‘˜,0 Initi nitializ ize weights ๐‘ฅ ๐‘—๐‘˜,0 = ๐บ(๐‘ก ๐‘—๐‘˜,๐‘ˆ ) Initi nitializ For ๐‘š ๐‘ž minimization, ๐บ ๐‘ฆ = ๐‘ฆ ๐‘žโˆ’2

Recommend


More recommend