Efficient computation of Robust Low-Rank Matrix Approximations in the Presence of Missing Data using the L 1 Norm Anders Eriksson and Anton van den Hengel CVPR 2010
Usual low rank approximation using L 2 norm– SVD. Robust low rank approximation using L 2 norm- Wiberg Algorithm. “Robust” low rank approximation in the presence of: – missing data –Outliers –L 1 norm –Generalization of Wiberg Algorithm. Y= UV MXN MXR RXN
Problem W is the indicator matrix, w ij = 1 if y ij is known, else 0.
Wiberg Algorithm W matrix indicates the presence/absence of elements From: “On the Wiberg algorithm for matrix factorization in the presence of missing components”, Okatani et al, IJCV 2006,
Alternating Least Squares To find the minimum of φ, find derivatives Considering the two equations independently. Starting with some initial estimates u 0 and v 0, update u from v and v from u. Converges very slowly, specially for missing components and strong noise. From: “On the Wiberg algorithm for matrix factorization in the presence of missing components”, Okatani et al, IJCV 2006,
Back to Wiberg In non-linear least squares problems with multiple parameters, when assuming part of the parameters to be fixed, minimization of the least squares with respect to the rest of the parameters becomes a simple problem, e.g., a linear problem. So closed form solutions may be found. Wiberg applied it to this problem of factorization of matrix with missing components. From: “On the Wiberg algorithm for matrix factorization in the presence of missing components”, Okatani et al, IJCV 2006,
Back to Wiberg For a fixed u, the L 2 norm becomes a linear, least squares minimization problem in v. –Compute optimal v*(u) Apply Gauss-Newton method to the above non-linear least squares problem to find optimal u*. Easy to compute derivative because of L 2 norm From: “On the Wiberg algorithm for matrix factorization in the presence of missing components”, Okatani et al, IJCV 2006,
Linear Programming and Definitions
L 1 -Wiberg Algorithm Minimization problem in terms of L 1 norm Minimization problem in terms of v and u independently Substituting v* into u
Comparing to L 2 -Wiberg V*(U) is not easily differentiable The minimization function (u,v*) is not a least squares minimization problem, so Gauss-Newton can’t be applied directly. Idea: Let V*(U) denote the optimal basic solution. V*(U) is differentiable assuming problem is feasible, as per Fundamental Theorem of differentiability of linear programs. Jacobian for the G-N :: derivative of solution to a linear prog. problem
≈ Add an additional term to the function and minimize the value of the term ?
Results Tested on synthetic data. –Randomly created measurement matrices Y drawn from a uniform distribution [-1,1]. –20% missing, 10% noise [-5,5]. Real data –Dinosaur sequence from oxford-vgg.
Structure from motion Projections of 319 points tracked over 36 views. Addition of noise to 10% points. Full 3d reconstruction ~ low rank matrix approximation. Above-residual for the visible points. In L 2 norm, reconstruction error is evenly distributed among all elements of residual. In L 1 norm, error concentrated on few elements.