Robust PCA and Robust Subspace Tracking: A Comparative Evaluation

Abstract

This paper provides a comparative theoretical and experimental evaluation of solutions for robust PCA and robust subspace tracking (dynamic RPCA) that rely on the sparse+lowrank matrix decomposition formulation. The emphasis is on simple and provably correct methods. Experimental comparisons are shown for video layering (separate a given video into foreground and background layer videos) which is a key first step in simplifying many video analytics and computer vision tasks, e.g., video denoising or activity recognition.