Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 X � � Φ T i ( x − x 0 ) � R ( x ) = λ i i x − x 0 λ 1 Φ i λ i ... � Λ = 0 x − � x T Φ i λ n i
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 X � � Φ T i ( x − x 0 ) � R ( x ) = λ i i x − x 0 λ 1 Φ i λ i ... � Λ = 0 x − � x T Φ i λ n i Φ 0
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 X � Φ T � i ( x − x 0 ) � R ( x ) = λ i i x − x 0 λ 1 Φ i λ i ... � Λ = 0 x − � x T Φ i λ n i Φ 0 Φ i
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 X � Φ T � i ( x − x 0 ) � R ( x ) = λ i i x − x 0 λ 1 Φ i λ i ... � Λ = 0 x − � x T Φ i λ n i Φ 0 Φ i Φ n − 1
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 X � Φ T � i ( x − x 0 ) � R ( x ) = λ i i x − x 0 λ 1 Φ i λ i ... � Λ = 0 x − � x T Φ i λ n i Φ 0 Φ i Φ n − 1 Φ n
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � � x − x 0 � R ( x ) = L
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 Σ ∝ L − 1
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 Σ ∝ L − 1 = ΦΛ − 1 Φ T
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 Σ ∝ L − 1 = ΦΛ − 1 Φ T = ΨΩΨ T
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � x − x 0 � � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 Σ ∝ L − 1 = ΦΛ − 1 Φ T = ΨΩΨ T σ 1 ... Ω = σ i σ n i
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � � x − x 0 � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 X Σ ∝ L − 1 = ΦΛ − 1 Φ T = ΨΩΨ T σ i Ψ i Ψ T = i i σ 1 ... Ω = σ i σ n i
Laplacian Regularisation x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 � 2 � � x − x 0 � R ( x ) = L ⇢ � � 1 Gaussian Distribution: 2 k x � µ k 2 N ( x ) / exp Σ − 1 µ = x 0 X Σ ∝ L − 1 = ΦΛ − 1 Φ T = ΨΩΨ T σ i Ψ i Ψ T = i i Ψ 0 σ 1 Ψ 1 ... Ω = σ i Ψ n σ n i
Example: Optical Flow ∆ x ||I 0 ( x ) + ∂ I 0 ( x ) ∆ x − I 1 ( x ) || 2 arg min 2 ∂ x T ∆ x 1 . ∆ x = . . ∆ x N N = no. of pixels
General Topology x 0 6 = What if grid?
General Topology x 0 6 = What if grid?
General Topology x 0 6 = What if grid? X X [( x i − x 0 i ) − ( x j − x 0 j )] 2 = R ( x ) j ∈ N i i
General Topology x 0 6 = What if grid? X X [( x i − x 0 i ) − ( x j − x 0 j )] 2 = R ( x ) j ∈ N i i x = x 0 + Ψ α
Smooth Deformation Basis ~Frequency x y z Ψ 0 Ψ 1 Ψ 2 Ψ 3 Ψ 25 Ψ 50
Smooth Deformation Basis ~Frequency x y z Ψ 0 Ψ 1 Ψ 2 Ψ 3 Ψ 25 Ψ 50
Heuristic Regularisation: Recap - Regularisation is important because image measurements are not enough - Priors model the space of valid instances of object’s geometry - Regularisers penalise object geometry outside the space of valid instances. - The smoothness the assumption is a good heuristic - Laplacian regularisers enforce smoothness by penalising high frequency variations more heavily than lower frequency variations - The concept of frequency that is penalised can be specialised to the topology of the object though defining specialised graph- laplacian - But... is that the best we can do? Prior Regulariser
Today • Parts Based Registration • Regularizing Parts (Heuristic) • Regularizing Parts (Learned) 30
Data Driven (Learned) Regularisers What if we have annotated data? [3] Huang et al.’07
Topology of Samples vs. Parts Parts Samples Data
Topology of Samples vs. Parts Parts < 3 d Samples Data Sample Topology
Topology of Samples vs. Parts Parts < d < 3 d Samples Data Sample Topology Part Topology
Recommend
More recommend