1. What Energy Functions Can be Minimized Using Graph Cuts? Shai Bagon Advanced Topics in Computer Vision June 2010

2. What is an Energy Function? E a number suggested solution For a given problem:Image Segmentation: 237-20 Useful Energy function: 1.Good solution  Low energy 2.Tractable  Can be minimized

3. Families of Functions or Outline F 2 submodular Non submodular F 3 Beyond F 3

4. Foreground Selection Let y i – color of i th pixel x i {0,1} BG/FG labels (variables) Given BG/FG scribbles: Pr(x i |y i ) =How likely each pixel to be FG/BG Pr(x m |x n ) =Adjacent pixels should have same label F 2 energy: E(x)=∑ i E i (x i )+∑ ij E ij (x i,x j ) xmxm xnxn xixi yiyi

5. Submodular Known concept from set-functions: E(x) = ∑ i E i (x i ) + ∑ ij E ij (x i, x j ), x i {0,1} 1CD 0AB xjxixjxi 01 E ij (x i,x j ): What does it mean? B+C-A-D ≥ 0

6. How to Minimize? E(x) = ∑ i E i (x i ) + ∑ ij E ij (x i, x j ), x i {0,1} Local “beliefs”: Data term Prior knowledge: Smoothness term F 2 submodular

7. Graph Partitioning A weighted graph G=( V E w ) Special Nodes: s t s-t cut: Cost of a cut: Nice property: 1:1 mapping s-t cut ↔ {0,1} |V|-2 w V E w ij s t    TjSi TSCut, ),( VTSTS TtSsVTVS  ,,,, 

8. s t Graph Partitioning - Energy E(x) = ∑ i E i (x i ) + ∑ ij E ij (x i, x j ) Graph Partitioning i j E j (1) D-C B+C-A-D E i (0) 1CD 0AB xjxixjxi 01 E ij (x i,x j ) C-A 00 D-C0 0 00 B+C-A-D0 = A +++ C-A

9. s t Graph Partitioning - Energy E(x) = ∑ i E i (x i ) + ∑ ij E ij (x i, x j ) Graph Partitioning i j E j (1) B+C-A-D E i (0) C-A D-C st cut  binary assignment cut cost  energy of assignment min cut  Energy min. B=E ij (0,1)

10. Recap F 2 submodular: E(x) = ∑ i E i (x i ) + ∑ ij E ij (x i, x j ) E ij (1,0)+E ij (0,1)≥E ij (0,0)+E ij (1,1) Mapping from energy to graph partition Min Energy = computing min-cut Global optimum in poly time for submodular functions!

11. Next… Multi-label F 2 E(x)=∑ i E i (x i ) + ∑ ij E ij (x i,x j ) s.t. x i {1,…,L} –Fusion moves: solving binary sub-problems –Applications to stereo, stitching, segmentation… ● Current labeling suggested labeling “Alpha expansion” = Fusion Solve Binary problem: x i =0 x i =1

12. Stereo matching see http://vision.middlebury.edu/stereo/http://vision.middlebury.edu/stereo/ Ground truth Pairwise MRF [Boykov et al. ‘01] slide by Carsten Rother, ICCV’09 Input:

13. Panoramic stitching slide by Carsten Rother, ICCV’09

14. Panoramic stitching slide by Pushmeet Kohli, ICCV’09

15. AutoCollage http://research.microsoft.com/en-us/um/cambridge/projects/autocollage/ [Rother et. al. Siggraph ‘05 ]

16. Next… Multi-label F 2 E(x)=∑ i E i (x i ) + ∑ ij E ij (x i,x j ) s.t. x i {1,…,L} –Fusion moves: solving binary sub-problems –Applications to stereo, stitching, segmentation… Non-submodular Beyond pair-wise interactions: F 3

17. Merging Regions input image regions (Ncuts) “edge” prob. pipi “weak” edge “strong” edge p i – prob. of boundary being edge GOAL: Find labeling x i {0,1} that max: i j min: Taking -log

18. Merging Regions Adding and subtracting the same number

19. Merging Regions Solving for edges: Consistency constraints: No “dangling” edge J x1x1 x2x2 x3x3 EJEJ 0000 1110 0110 001 λ wiwi xixi No longer pair-wise: F 3

20. Minimization trick Freedman D., Turek MW, Graph cuts with many pixel interactions: theory and applications to shape modeling. Image Vision Computing 2010

21. Merging Regions The resulting energy: + Pair-wise - Non submodular!

22. Quadratic Pseudo-Boolean Optimization s i j t ij Kolmogorov V., Carsten R., Minimizing non-submodular functions with graph cuts – a review. PAMI ’ 07

23. + All edges with positive capacities - No constraint Labeling rule: partial labeling s i j t ij Quadratic Pseudo-Boolean Optimization

24. Properties of partial labeling y: 1. Let z=FUSE(y,x)  E(z)≤E(x) 2. y is subset of optimal y* y is complete: 1. E submodular 2. Exists flipping (inference in trees) s i j t ij Quadratic Pseudo-Boolean Optimization

25. 0????? rpqst 000?? 0010? rpqst rpqst QPBO: Probe Node p: 0 1 What can we say about variables? r -> is always 0 s -> is always equal to q t -> is 0 when q = 1 slide by Pushmeet Kohli, ICCV’09 QBPO - Probing

26. Probe nodes in an order until energy unchanged Simplified energy preserves global optimality and (sometimes) gives the global minimum slide by Pushmeet Kohli, ICCV’09 QBPO - Probing

27. Merging Regions Result using QPBO-P: Result regions (Ncuts)input image

28. Recap F 3 and more –Minimization trick Non submodular –QPBO approx. – partial labeling

29. Beyond F 3 … [Kohli et. al. CVPR ‘07, ‘08, PAMI ’08, IJCV ‘09]

30. Image Segmentation E(X) = ∑ c i x i + ∑ d ij |x i -x j | ii,j E: {0,1} n → R 0 → fg, 1 → bg n = number of pixels [Boykov and Jolly ‘ 01] [Blake et al. ‘04] [Rother et al.`04] Image Unary Cost Segmentation

31. P n Potts Potentials Patch Dictionary (Tree) C max  0 { 0 if x i = 0, i p C max otherwise h(X p ) = p [slide credits: Kohli]

32. P n Potts Potentials E(X) = ∑ c i x i + ∑ d ij |x i -x j | + ∑ h p (X p ) ii,j p p { 0 if x i = 0, i p C max otherwise h(X p ) = E: {0,1} n → R 0 → fg, 1 → bg n = number of pixels [slide credits: Kohli]

33. Image Segmentation E(X) = ∑ c i x i + ∑ d ij |x i -x j | + ∑ h p (X p ) ii,j ImagePairwise SegmentationFinal Segmentation p E: {0,1} n → R 0 → fg, 1 → bg n = number of pixels [slide credits: Kohli]

34. Application: Recognition and Segmentation from [Kohli et al. ‘08] Image Unaries only TextonBoost [Shotton et al. ‘06] Pairwise CRF only [Shotton et al. ‘06] P n Potts One super- pixelization another super- pixelization

35. Robust(soft) P n Potts model { 0 if x i = 0, i p f( ∑ x p ) otherwise h(x p ) = p p from [Kohli et al. ‘08] Robust P n PottsP n Potts

36. Application: Recognition and Segmentation From [Kohli et al. ‘08] Image Unaries only TextonBoost [Shotton et al. ‘06] Pairwise CRF only [Shotton et al. ‘06] P n Potts robust P n Potts (different f) One super- pixelization another super- pixelization

37. Same idea for surface-based stereo [Bleyer ‘10] One input image Ground truth depth Stereo with hard-segmentation Stereo with robust P n Potts This approach gets best result on Middlebury Teddy image-pair:

38. How is it done… H (X) = F ( ∑ x i ) Most general binary function: H (X) ∑ x i concave 0 The transformation is to a submodular pair-wise MRF, hence optimization globally optimal [slide credits: Kohli]

39. Higher order to Quadratic Start with P n Potts model: { 0 if all x i = 0 C 1 otherwise f(x) = x {0,1} n min f(x) min C 1 a + C 1 (1-a) ∑ x i x = x,a {0,1} Higher Order Function Quadratic Submodular Function ∑ x i = 0 a=0 f(x) = 0 ∑ x i > 0 a=1f(x) = C 1 [slide credits: Kohli]

40. Higher order to Quadratic min f(x) min C 1 a + C 1 (1-a) ∑ x i x = x,a {0,1} Higher Order FunctionQuadratic Submodular Function ∑xi∑xi 1 23 C1C1 C1∑xiC1∑xi [slide credits: Kohli]

41. Higher order to Quadratic min f(x) min C 1 a + C 1 (1-a) ∑ x i x = x,a {0,1} Higher Order Submodular Function Quadratic Submodular Function ∑xi∑xi 1 23 C1C1 C1∑xiC1∑xi a=1 a=0 Lower envelope of concave functions is concave [slide credits: Kohli]

42. Summary Submodular F 2 F 3 and beyond: minimization trick Non submodular –QPBO(P) Beyond F 3 – Robust HOP s i j t ij ∑xi∑xi a=1 a=0 f 2 (x) f 1 (x)