Auxiliary Cuts for General Classes of Higher-Order Functionals 1 Ismail Ben Ayed, Lena Gorelick and Yuri Boykov
Standard Segmentation Functionals 2 S
Historic Data Linear terms are not enough 3 Standard model Learned distributions
Linear terms are not enough 3 Segmentation with log likelihoods Learned distributions
Linear terms are not enough 3 Standard model Target distributions Segmentation with log likelihoods
Linear terms are not enough 3 Standard model Target distributions
Segmentation with log likelihoods Linear terms are not enough 3 Standard model Target distributions
Segmentation with log likelihoods Linear terms are not enough 3 Standard model Learned distributions Obtained distributions
From log-likelihoods to higher-order terms 4 Rother et al. 06, Ben Ayed CVPR 10, Gorelick et al. ECCV 12, Jiang et al. CVPR 12
Standard vs. High-order 5 Input High-order Likelihoods High-order Input
Regional Functional Examples Volume Constraint 6
Bin Count Constraint Regional Functional Examples Volume Constraint 6
7 Contribution: Bound Optimization of General Higher-Order Terms Non-Linear Combination of Linear Terms
Optimization Higher-order Pairwise Sub-modular 8
9 Prior Art: General-Purpose Techniques Based on Functional Derivatives
9 -- Level Sets: Ben Ayed et al. CVPR Line search: Gorelick et al. ECCV 2012 Can be slow
9 -- Level Sets: Ben Ayed et al. CVPR Line search: Gorelick et al. ECCV 2012 F differentiable Prior Art: General-Purpose Techniques Based on Functional Derivatives
9 -- Level Sets: Ben Ayed et al. CVPR Line search: Gorelick et al. ECCV 2012 Parameters? Prior Art: General-Purpose Techniques Based on Functional Derivatives
Prior Art: Specialized Techniques 9 Volume constraint: Werner, CVPR 2008 Norms between bin counts: Mukherjee et al. CVPR 2009, Jiang et al. CVPR 2012 Bhattacharyya : Ben Ayed et al. CVPR 2010, Punithakumar et al. SIAM 2012 Only particular cases
Auxiliary Function Optimization 10
Auxiliary Function Optimization 10
Auxiliary Function Optimization 10
Standard Tricks for Deriving Auxiliary Functions 12 Cauchy-Schwarz inequality Quadratic bound principle First-order expansion Jensen’s inequality E.g.: EM is based on this approach
Jensen’s Inequality bound 11
Unary Terms Jensen’s Inequality bound
11 Jensen’s Inequality bound
Auxiliary Function Derivation 13
Auxiliary Function Derivation 13
Auxiliary Function Derivation 13
Auxiliary Function Derivation 13 Constant
Auxiliary Function Derivation 13 Sum to 1
Auxiliary Function Derivation 13 Jensen’sLinear auxiliary function
Difference with other methods: the volume constraint case 14
Difference with other methods: the volume constraint case 14 Gradient Descent
Difference with other methods: the volume constraint case 14 Trust Region: Gorelick et al. CVPR 13
Difference with other methods: the volume constraint case 14 Auxiliary Cuts
General Form of the Functionals 15 Higher-order Sub-modular
General Form of the Functionals 15 Linear bound Sub-modular Higher-order Sub-modular
General Form of the Functionals 15 Graph Cut Higher-order Sub-modular Linear bound Sub-modular
Experimental examples
L2 Bin Count (Aux. Cuts vs. Level Sets) Level-Set, dt=1 Level-Set, dt=50 Level-Set, dt=1000 Init Aux. Cuts 16
User input Result User input Iter 2 User input ResultB-J Initial segment Iter. 3Iter L1 Bin Count
18 inputs Input L2 Volume Constraint User input B-J B-J and Volume
Conclusions 19 Advantages: Derivative-free No optimization parameters, e.g., step size Easy to implement Never worsen the energy at each iteration
Conclusions 19 Limitations: The form of F should verify some conditions Limited to nested evolutions of segments
Conclusions 19 Extensions: More general forms of F Arbitrary evolutions of segments
19 inputs Input Thanks