Robust Alignment by Sparse and Low-rank Decomposition for Linearly Correlated Images Zhihui Zhu Instructor: Prof. Hoff Colorado School of Mines December 2, 2013 1 Outlier • Introduction • Previous work on alignment • Alignment by sparse and low rank decomposition • Experiment results • Conclusion 2 Introduction • Multiple images of an object • Challenging existing vision alg.: face detection etc. • Because of: illumination variation, partial occlusion, no alignment 3 Recognition Pipeline Detection Alignment Identification Recognition Pipeline[1]: Hw 2 Alignment: using affine/geometric transformation to match 4 Batch Image Alignment • Given many images of an object, simultaneously align them to a fixed canonical template • Mathematically: Given Recover observations {I i } images {I i0 } = I i ( I i0 + ei ) τ i−1 transformaitons {τ } −1 i 5 Outlier • Introduction • Previous work on alignment • Alignment by sparse and low rank decomposition • Experiment results • Conclusion 6 Previous work • Learned-Miller’s congealing alg.[2]: minimize the sum of entropies • Least squares congealing alg.[3]: Minimize the sum of squared distances Rank 1 Low rank • Minimize log-determinant measure [4] 7 Outlier • Introduction • Previous work on alignment • Alignment by sparse and low rank decomposition • Experiment results • Conclusion 8 Modeling Misalignments as Domain Transformations • I1 and I 2 misaligned images, exists τ ∈ Aff(2) = I 2 ( x , y ) (= I1 τ )( x, y ) I1 (τ ( x, y )) • 2-D Affine group (Aff(2)): G 0 t 1 Hw 2 x g11 g12 t x u y = g g t v 21 22 y 1 0 0 1 1 1.23 0.49 -55.14 τ = −0.11 0.87 23.19 0 0 1 9 Matrix Rank as Measure of Image Similarity • Well-aligned images are linearly correlated 0 0 m×n = A vec I ⋅⋅⋅ vec I ∈ R [ ( ) | | ( )] • Matrix 1 n should be row rank If {I i0 } are the same rank one 0 { I If i } are images of convex Lambertian object under varying illumination rank 9 approx. [5] Fig. The same face, under two different lighting conditions[5]. 10 Modeling Corruption as Large, Sparse Errors • Corruptions or occlusions break low-rank structure • Fortunately, affect only a small fraction model as sparse errors 11 Robust Alignment of Correlated Images • Given • Recover observations {I i } images {I i0 } = I i ( I i0 + ei ) τ i−1 transformaitons {τ } −1 i min rank ( A) + γ || E ||0 s.t. D τ = A+ E A, E ,τ = D τ [vec( I1 τ 1 ) | ⋅⋅⋅ | vec( I n τ n )] 12 Convex Relaxation NP-hard • Rank(A) • l0 -norm Relaxation method • Nuclear norm || A ||∗ = ∑ σ i ( A) • l1 -norm || A ||1 i 13 Iterative Linearization • Approximate by Taylor Series ∂ = ;J f (x + ∆ x) f ( x) + J ⋅ (∆x) + h.o.t= f (ξ ) |ξ = x ∂ξ • Here ∂ I (τ + ∆τ ) ≈ I τ + J ∆τ ; J = ( I ξ ) |ξ =τ ∂ξ • And D (τ + ∆τ ) ≈ D τ + ∑ J i ∆τ i eiT i ∂ T = Ji = ( Ii ξ ) |ξ =τ i , ei [0 0 1 0 0] ∂ξ 14 Iterative Convex Programming • Leads to a convex optimization problem: T A+ E min || A ||∗ +γ || E ||1 s.t. D τ + ∑ J i ∆τ i ei = A, E , ∆τ i Main Iteration Initialization k =0 τ 1 , ,τ n k= k + 1 1. Compute Jocobian : J i 2. Warp and Normalize : D τ = [ vec ( I 1 τ 1 ) || vec ( I 1 τ 1 ) ||2 || vec ( I n τ n ) || vec ( I n τ n ) ||2 ] 3. Solve above linearized convex optimization : Accelerated Proximal Gradient (APG) 4. Update Transforamtion : τ = τ + ∆τ 4. No ∆τ 2 ≤ ε Yes Stop 15 Outlier • Introduction • Previous work on alignment • Alignment by sparse and low rank decomposition • Experiment results • Conclusion 16 Alignment Example Original images Aligned images 17 Alignment Example Low rank part A average of unaligned D Sparse errors E average of aligned D average of A 18 Comparison Rotation [ −10 ,10 ] x- and y-translations: [ −3, 3] pixels Occluded by 12 x12 batch (a) Original perturbed and corrupted images (b) Alignment results by [4] (c) Alignment results by RASL 19 Comparison (d) Alignment results by RASL: upper: low rank part; down: sparse errors Mean error Error std. Max error Initial misalignment 2.5 1.03 4.87 [23] 1.66 0.85 4.02 RASL 0.48 0.23 1.07 20 Removing occlusions Original homograph images Aligned images 21 Removing occlusions Reconstructed images A average of unaligned D Removed occlusions E average of aligned D average of A 22 Outlier • Introduction • Previous work on alignment • Alignment by sparse and low rank decomposition • Experiment results • Conclusion 23 Conclusion • Seeks a set of image domain transformations • Low rank + sparse structure seems a good model for aligned images • Experiments verify this point • Extension to large scale situations. • Online robust image alignment? 24 25 Reference [1] Gary B. Huang, Vidit Jain, and Erik Learned-Miller. Unsupervised joint alignment of complex images. ICCV, 2007. [2] E. Learned-Miller. Data driven image models through continuous joint alignment. PAMI, 2006. [3] M. Cox, S. Lucey, S. Sridharan, and J. Cohn. Least squares congealing for unsupervised alignment of images. CVPR, 2008. [4] A. Vedaldi, G. Guidi, and S. Soatto. Joint alignment up to (lossy) transforamtions. CVPR, 2008. [5] R. Basri and D. Jacobs. Lambertian reflectance and linear subspaces. PAMI, 2003. [6] http://perception.csl.illinois.edu/matrix-rank/rasl.html 26