Frank Hai Ong (Berkeley)
Wednesday, 15.3.2017, 11:30
Data matrices constructed from multimedia data are often correlated at different scales. Motivated by this observation, we consider the decomposition of a matrix into block-wise low rank components of multiple scales. We approach the problem via a convex formulation and present an iterative algorithm using block-wise SVD’s. We show that in practice, the multi-scale low rank decomposition often returns intuitive matrix decomposition. We also show results on real-world applications, including shadow removal of face images, background subtraction in surveillance videos, dynamic magnetic resonance image reconstruction and movie rating matrix completion.