Title: Ordinary Least-Squares
1Regularized Least-Squares
2Outline
- Why regularization?
- Truncated Singular Value Decomposition
- Damped least-squares
- Quadratic constraints
3Why regularization?
4Why regularization?
- We have seen that
- But what happens if the system is almost
dependent? - The solution becomes very sensitive to the data
- Poor conditioning
5The 1-dimensional case
- The 1-dimensional normal equation
6The 1-dimensional case
- The 1-dimensional normal equation
7The 1-dimensional case
- The 1-dimensional normal equation
8Why regularization
- Contradiction between data and model
9A more interesting examplescattered data
interpolation
10True curve
11Radial basis functions
12Radial basis functions
13Rbf are popular
- Modeling
- J. C. Carr, R. K. Beatson, J. B. Cherrie, T. J.
Mitchell,W. R. Fright, B. C. McCallum, and T. R.
Evans. Reconstruction and representation of 3d
objects with radial basis functions. In
Proceedings of ACM SIGGRAPH 2001, Computer
Graphics Proceedings, Annual Conference Series,
pages 6776, August 2001. - G. Turk and J. F. OBrien. Modelling with
implicit surfaces that interpolate. ACM
Transactions on Graphics, 21(4)855873, October
2002. - Animation
- J. Noh and U. Neumann. Expression cloning. In
Proceedings of ACMSIGGRAPH 2001, Computer
Graphics Proceedings, Annual Conference Series,
pages 277288, August 2001. - F. Pighin, J. Hecker, D. Lischinski, R. Szeliski,
and D. H. Salesin. Synthesizing realistic facial
expressions from photographs. In Proceedings of
SIGGRAPH 98, Computer Graphics Proceedings,
Annual Conference Series, pages 7584, July 1998.
14Radial basis functions
15Radial basis functions
- At every point
- Solve the least-squares problem
16Radial basis functions
- At every point
- Solve the least-squares problem
17Rbf results
18pi0 close to pi1
19Radial basis functions
- At every point
- Solve the least-squares problem
20Radial basis functions
- At every point
- Solve the least-squares problem
- If pi0 close to pi1, A is near singular
21pi0 close to pi1
22pi0 close to pi1
23Rbf results with noise
24Rbf results with noise
25The Singular Value Decomposition
- Every matrix A (nxm) can be decomposed into
- where
- U is an nxn orthogonal matrix
- V is an mxm orthogonal matrix
- D is an nxm diagonal matrix
26The Singular Value Decomposition
- Every matrix A (nxm) can be decomposed into
- where
- U is an nxn orthogonal matrix
- V is an mxm orthogonal matrix
- D is an nxm diagonal matrix
27Geometric interpretation
28Solving with the SVD
29Solving with the SVD
30Solving with the SVD
31Solving with the SVD
32Solving with the SVD
33A is nearly rank defficient
34A is nearly rank defficient
35A is nearly rank defficient
36A is nearly rank defficient
- A is nearly rank defficient
- gtsome of the are close to 0
37A is nearly rank defficient
- A is nearly rank defficient
- gtsome of the are close to 0
- gtsome of the are close to
38A is nearly rank defficient
- A is nearly rank defficient
- gtsome of the are close to 0
- gtsome of the are close to
- Problem with
39A is nearly rank defficient
- A is nearly rank defficient
- gtsome of the are close to 0
- gtsome of the are close to
- Problem with
- Truncate the SVD
40pi0 close to pi1
41Rbf fit with truncated SVD
42Rbf results with noise
43Rbf fit with truncated SVD
44Choosing cutoff value k
45Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
46Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
?
47Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
48Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
- Skinning
- Inverse skinning
- Let be a set of pairs of geometry and skeleton
configurations
49Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
- Skinning
- Inverse skinning
- Let be a set of pairs of geometry and skeleton
configurations
50Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
- Skinning
- Inverse skinning
- Let be a set of pairs of geometry and skeleton
configurations
51Example inverse skinningSkinning Mesh
Animations, James and Twigg, siggraph
- Skinning
- Inverse skinning
- Let be a set of pairs of geometry and skeleton
configurations
52Skinning Mesh Animations, James and Twigg,
siggraph
53Problem with the TSVD
- We have to compute the SVD of A, and O() process
- impractical for large marices
- Little control over regularization
54Damped least-squares
- Replace
- by
- where is a scalar and L is a matrix
55Damped least-squares
- Replace
- by
- where is a scalar and L is a matrix
- The solution verifies
56Examples of L
Diagonal
Differential
Limit scale
Enforce smoothness
57Rbf results with noise
58(No Transcript)
59(No Transcript)
60(No Transcript)
61Example Least-Squares Meshes, Sorkin and
Cohen-Or, siggaph
- Reconstruct a mesh given
- Control points
- Connectivity (planar mesh)
62Example Least-Squares Meshes, Sorkin and
Cohen-Or, siggaph
- Reconstruct a mesh given
- Control points
- Connectivity (planar mesh)
- Smooth reconstruction
63Example Least-Squares Meshes, Sorkin and
Cohen-Or, siggaph
- Reconstruct a mesh given
- Control points
- Connectivity (planar mesh)
- Smooth reconstruction
- In matrix form
64Reconstruction
- Minimize reconstruction error
- where
65Least-Squares Meshes, Sorkine and Cohen-Or,
siggraph
66Quadratic constraints
67Quadratic constraints
68Example
69Example
70Example
71Discussion
- If , there is no
solution (since there is no x for which
)
72Discussion
- If , there is no
solution (since there is no x for which
) - If , the solution exists
and is unique
73Discussion
- If , there is no
solution (since there is no x for which
) - If , the solution exists
and is unique - Either the solution of
is in the feasible set
74Discussion
- If , there is no
solution (since there is no x for which
) - If , the solution exists
and is unique - Either the solution of
is in the feasible set - or the solution is at the boundary
- Solve
75Discussion
- Solve
- where is a Lagrange multiplier
76Conclusion
- TSVD really useful if you need an SVD
77Conclusion
- TSVD really useful if you need an SVD
- Regularization constrains the solution
- Value, differential operator, other properties
- Soft (damping) or hard constraint (quadratic)
- Linear or non-linear
78Conclusion
- TSVD really useful if you need an SVD
- Regularization constrains the solution
- Value, differential operator, other properties
- Soft (damping) or hard constraint (quadratic)
- Linear or non-linear
- Danger of over-damping or constraining
79Example inverse kinematic
- Problem solve for joint angles given
end-effector positions
80Example inverse kinematic
- Problem solve for joint angles given
end-effector positions
?
81Example inverse kinematic
- Problem solve for joint angles given
end-effector positions
82Example inverse kinematic
- Problem solve for joint angles given
end-effector positions