1. C. Olsson Higher-order and/or non-submodular optimization: Yuri Boykov jointly with Western University Canada O. Veksler Andrew Delong L. Gorelick C. NieuwenhuisE. Toppe I. Ben Ayed A. Delong H. IsackA. Osokin M. Tang Tutorial on Medical Image Segmentation: Beyond Level-Sets MICCAI, 2014

2. Basic segmentation energy E(S) boundary smoothness (quadratic / pairwise) segment appearance (linear / unary) - such second-order functions can be minimized exactly via graph cuts [Greig et al.’91, Sullivan’94, Boykov-Jolly’01] n-links s t a cut t-link

3. Basic segmentation energy E(S) boundary smoothness (quadratic / pairwise) segment appearance (linear / unary) - submodular second-order functions can be minimized exactly via graph cuts [Greig et al.’91, Sullivan’94, Boykov-Jolly’01] n-links s t a cut t-link [Hammer 1968, Pickard&Ratliff 1973]

4. any (binary) segmentation functional E(S) is a set function E: S Ω Submodular set functions

5. Set function is submodular if for any Significance: any submodular set function can be globally optimized in polynomial time [Grotschel et al.1981,88, Schrijver 2000] S T Ω Edmonds 1970 (for arbitrary lattices)

6. Submodular set functions Set function is submodular if for any S T Ω equivalent intuitive interpretation via “diminishing returns” v Easily follows from the previous definition: Significance: any submodular set function can be globally optimized in polynomial time [Grotschel et al.1981,88, Schrijver 2000]

7. Assume set Ω and 2nd-order (quadratic) function Function E(S) is submodular if for any Significance: submodular 2 nd -order boolean (set) function can be globally optimized in polynomial time by graph cuts [Hammer 1968, Pickard&Ratliff 1973] Indicator variables [Boros&Hammer 2000, Kolmogorov&Zabih2003] Submodular set functions

8. Combinatorial optimization Continuous optimization Global Optimization submodularityconvexity ?

9. Global Optimization f (x,y) =  ∙ xy for  0 1 1 y x A C D B EXAMPLE: 0 1 1 y x A C D B f (0,0) + f (1,1) ≤ f (0,1) + f (1,0) submodular binary energy convex continuous extension submodularityconvexity ?

10. submodularityconvexity Global Optimization 0 1 1 y x A C D B f (0,0) + f (1,1) ≤ f (0,1) + f (1,0) 0 1 1 y x A C D B submodular binary energy concave continuous extension f (x,y) =  ∙ xy for  EXAMPLE:

11. Assume Gibbs distribution over binary random variables for posterior optimization (in Markov Random Fields) Theorem [Boykov, Delong, Kolmogorov, Veksler in unpublished book 2014?] All random variables s p are positively correlated iff set function E(S) is submodular That is,… submodularity implies MRF with “smoothness” prior

12. second-order submodular energy boundary smoothness segment region/appearance w pq < 0 would imply non-submodularity

13. Now - more difficult energies Non-submodular energies High-order energies

14. Beyond submodularity: even second-order is challenging Example: deconvolution image I blurred with mean kernel QPBO TRWS “partial enumeration” submodular quadratic term non-submodular quadratic term

15. Beyond submodularity: even second-order is challenging Example: quadratic volumetric prior e.g. NOTE: any convex cardinality potentials are supermodular [Lovasz’83] (not submodular)

16. Many useful higher-order energies Cardinality potentials Curvature of the boundary Shape convexity Segment connectivity Appearance entropy, color consistency Distribution consistency High-order shape moments …

17. QPBO [survey Kolmogorov&Rother, 2007] LP relaxations [e.g. Schlezinger, Komodakis, Kolmogorov, Savchinsky,…] Message passing, e.g. TRWS [Kolmogorov] Partial Enumeration [Olsson&Boykov, 2013] Trust Region [e.g. Gorelick et al., 2013] Bound Optimization [Bilmes et al.’2006, Ben Ayed et al.’2013, Tang et al.’2014] Submodularization [e.g. Gorelick et al., 2014] Non-submodular and/or high-order functions E(S) Optimization is a very active area of research…

18. Our recent methods: local submodular approximations (submodularization) gradient descent + level sets (in continuous case) LP relaxations (QPBO, TRWS, etc) local linear approximations (parallel ICM) Prior approximation techniques based on linearization Fast Trust Region [CVPR 13] Auxiliary Cuts [CVPR 13] LSA [CVPR 14] Pseudo-bounds [ECCV 14], [Bilmes et al.2009]

19. Trust Region Approximation |S||S|  submodular terms appearance log-likelihoodsboundary length non-submodular term volume constraint Linear approx. at S 0 S0S0 S0S0 submodular approx. trust region L 2 distance to S 0

20. Volume Constraint for Vertebrae segmentation Log-Lik. + length 20

21. 1 Trust region Vs. bound optimization Fast iterative techniques Sub-problem solutions Convex continuous formulation Graph Cuts (submodular discrete formulation) Gradient descent and Level sets Vs.

22. Trust Region vs. Gradient Descent 2 iso-surfaces of approximation  E S E S - trust region ||S – S 0 || ≤ d gradient Gorelick et al., CVPR 2013

23. Trust Region vs. Gradient Descent 3 iso-surfaces of approximation  E S E S - trust region ||S – S 0 || ≤ d gradient e.g., Gorelick et al., CVPR 2013 Can be solved globally with graph cuts or convex relaxation

24. Bound Optimization solution space S StSt S t+1 At(S)At(S) E(S)E(S) E(St)E(St) E(S t+1 ) (Majorize-Minimize, Auxiliary Function, Surrogate Function) e.g., Ben Ayed et al., CVPR 2013 4

25. Bound Optimization solution space S StSt S t+1 S t+2 A t+1 (S) E(S)E(S) E(St)E(St) E(S t+1 ) E(S t+2 ) local minimum (Majorize-Minimize, Auxiliary Function, Surrogate Function) e.g., Ben Ayed et al., CVPR 2013 4

26. Bound Optimization solution space S StSt S t+1 S t+2 A t+1 (S) E(S)E(S) E(St)E(St) E(S t+1 ) E(S t+2 ) Can be solved globally with graph cuts or Convex relaxation (Majorize-Minimize, Auxiliary Function, Surrogate Function) e.g., Ben Ayed et al., CVPR 2013 4

27. Pseudo-Bound Optimization (Majorize-Minimize, Auxiliary Function, Surrogate Function) Tang et al., ECCV 2014 Make larger move! (a) (b) (c) solution space S StSt S t+1 E(S)E(S) E(St)E(St) E(S t+1 ) 4

28. 5 An example illustrating the difference between the three frameworks Initialization Adding volume Log-likelihood Gradient Descent

29. 5 An example illustrating the difference between the three frameworks Initialization Adding volume Log-likelihood Trust Region

30. 5 An example illustrating the difference between the three frameworks Initialization Adding volume Log-likelihood Bound Optimization

31. 5 An example illustrating the difference between the three frameworks Initialization Adding volume Log-likelihood Pseudo-Bound Optimization

32. 6 Summary of main differences/similarities FrameworkPropertiesApplicability Gradient Descent (e.g. level- sets) - Fixed small moves - Linear approximation Any differentiable functional Trust region - Adaptive large moves - Submodular or convex approximation Any differentiable functional Bound optimization - Unlimited large moves - Submodular or convex bounds Need a bound (not always easy)

33. 7 Trust Region Vs. Level sets: Experimental comparisons  Level set evolution without re initialization C. Li et al., CVPR 2005  No ad hoc initialization procedures  Keep a distance function by adding:  Frequently used by the community (>1500 citations)  Allows relatively large values of dt

34. High Low Min for a circle Compactness (Circularity) prior Trust region Vs. Level sets: Example 1 Ben Ayed et al., MICCAI 14 8

35. Trust region Compactness (Circularity prior) Trust region Vs. Level sets: Example 1 Without compactness 8

36. Trust region Vs. Level sets: Example 2 Volume constraint + Length Discrete Continuous 9

37. Trust Region Vs. Level Sets: Example 2 Volume Constraint + Length Init LS, t=1 LS, t=5 LS, t=10 LS, t=50 LS, t=1000 FTR, α=10FTR, α=5 FTR, α=2 9

38. Volume Constraint + Length Trust Region Vs. Level Sets: Example 2 10

39. Volume Constraint + Length Trust Region Vs. Level Sets: Example 2 10

40. Volume Constraint + Length Trust Region Vs. Level Sets: Example 2 10

41. Level-Set, t=50 Level-Set, t=1 Level-Set, t=5 Level-Set, t=10 Level-Set, t=1000 Init FTR, α=10 FTR, α=2 FTR, α=5 Trust Region Vs. Level Sets: Example 3 Shape moment Constraint + Length Up to order-2 moments learned from user-provided ellipse 11

42. Level-Set, t=50Level-Set, t=1Level-Set, t=5Level-Set, t=10 Level-Set, t=1000 Shape moment Constraint + Length Trust Region Vs. Level Sets: Example 3 11

43. Level-Set, t=50Level-Set, t=1Level-Set, t=5Level-Set, t=10 Level-Set, t=1000 Shape moment Constraint + Length Trust Region Vs. Level Sets: Example 3 11

44. Level-Set, t=1 Level-Set, t=1000 Shape moment Constraint + Length Trust Region Vs. Level Sets: Example 3 11

45. Level-Set, dt=1 Level-Set, dt=50 Level-Set, dt=1000 Init Surrogate 12 L2 bin count Constraint Bound Optimization Vs. Level Sets

46. L2 bin count Constraint Bound Optimization Vs. Level Sets 12

47. L2 bin count Constraint Bound Optimization Vs. Level Sets 12

48. Entropy-based segmentation Interactive segmentation with box volume balancing color consistency boundary smoothness |S||S| |Si||Si|   ii   i  ++ submodular termsnon-submodular term

49. 1.8 sec 1.3 sec 11.9 sec 576 sec Pseudo-Bound optimization [ECCV 2014] One-Cut [ICCV 2013] GrabCut (block-coordinate descent) Dual Decomposition Entropy-based segmentation

50. Curvature optimization instead of boundary length (pairwise submodular term) (3-rd order non-submodular potential, see next slide)

51. S S 3-cliques with configurations (0,1,0) and (1,0,1) p p+  p-  general intuition example Nieuwenhuis et al., CVPR 2014 more responses where curvature is higher

52. Need to penalize 3-clique configurations (010) and (101) uses submodular approximation and Trust Region [CVPR13, CVPR14] Nieuwenhuis et al., CVPR 2014

53. Segmentation Examples length-based regularization

54. elastica [Heber,Ranftl,Pock, 2012] Segmentation Examples

55. 90-degree curvature [El-Zehiry&Grady, 2010] Segmentation Examples

56. our squared curvature Segmentation Examples

57. Take-home messages - submodular or convex approximations are a good alternative to linearization methods (Level-Sets, QPBO, TRWS, etc.,) (Trust Region, Auxiliary Functions, Pseudo-bounds, etc.) [CVPR 2013, CVPR 2014, ECCV 2014] - code available