probably concave graph matching
play

(Probably) Concave Graph Matching Haggai Maron and Yaron Lipman - PowerPoint PPT Presentation

(Probably) Concave Graph Matching Haggai Maron and Yaron Lipman Weizmann Institute of Science Graph Matching n tr( ) min Graph Matching n tr( ) min DS Previous Work


  1. (Probably) Concave Graph Matching Haggai Maron and Yaron Lipman Weizmann Institute of Science

  2. Graph Matching π‘ŒβˆˆΞ  n βˆ’tr(π‘©π‘Œπ‘ͺπ‘Œ π‘ˆ ) min

  3. Graph Matching π‘ŒβˆˆΞ  n βˆ’tr(π‘©π‘Œπ‘ͺπ‘Œ π‘ˆ ) min DS

  4. Previous Work β€’ Superiority of the indefinite relaxation β€’ [Lyzinski et al. PAMI 2016] β€’ Efficient graph matching via concave energies β€’ [Vestner et al. CVPR 2017, Boyarski et al. 3DV 2017]

  5. Advantages of Concave Relaxations β€’ All local minima are permutation matrices πΉπ‘œπ‘“π‘ π‘•π‘§ πΈπ‘π‘›π‘π‘—π‘œ

  6. Many important graph matching problems are concave!

  7. Which 𝐡, 𝐢 give rise to concave relaxations?

  8. Concavity of Indefinite Relaxation β€’ Theorem: It is sufficient that 𝐡 = Ξ¦ 𝑦 𝑗 βˆ’ 𝑦 π‘˜ , 𝐢 = Ξ¨(𝑧 𝑗 βˆ’ 𝑧 π‘˜ ) where Ξ¦, Ξ¨ are positive definite functions of order one.

  9. Concavity of Indefinite Relaxation β€’ Theorem: It is sufficient that 𝐡 = Ξ¦ 𝑦 𝑗 βˆ’ 𝑦 π‘˜ , 𝐢 = Ξ¨(𝑧 𝑗 βˆ’ 𝑧 π‘˜ ) where Ξ¦, Ξ¨ are positive definite functions of order one.

  10. Concave Energies Euclidean distance in any dimension 𝐡 π‘—π‘˜ = ||𝑦 𝑗 βˆ’ 𝑦 π‘˜ || 2 β€’ Mahalanobis distances β€’ Spectral graph distances β€’ Matching objects with deep descriptors Spherical distance in any dimension 𝐡 π‘—π‘˜ = 𝑒 𝑇 π‘œ (𝑦 𝑗 , 𝑦 π‘˜ ) [bogomolny, 2007]

  11. Do we really need a concave relaxation?

  12. Do we really need a concave relaxation? Image taken from Crane et al. 2017

  13. Probably Concave Energies β€’ Theorem (upper bound on the probability of convex restriction) Let 𝑁 ∈ ℝ 𝑛×𝑛 and 𝐸 ≀ ℝ 𝑛 a uniformly sampled 𝑒 -dimensional subspace, then: π‘œ 1 βˆ’ 2π‘’πœ‡ 𝑗 βˆ’π‘’/2 𝑄𝑠 𝑁 𝐸 ≻ 0 ≀ min 𝑒 𝑗=1

  14. Probably Concave Energies β€’ Theorem (upper bound on the probability of convex restriction) Let 𝑁 ∈ ℝ 𝑛×𝑛 and 𝐸 ≀ ℝ 𝑛 a uniformly sampled 𝑒 -dimensional subspace, then: π‘œ 1 βˆ’ 2π‘’πœ‡ 𝑗 βˆ’π‘’/2 𝑄𝑠 𝑁 𝐸 ≻ 0 ≀ min 𝑒 𝑗=1

  15. Applications

  16. Conclusion β€’ A large family of concave or probably concave relaxations β€’ Checking probable concavity with eigenvalue bound β€’ Extension of [Lyzinsky et al. 2016] to practical matching problems

  17. The End β€’ Support β€’ ERC Grant (LiftMatch) β€’ Israel Science Foundation β€’ Thanks for listening!

Recommend


More recommend